M A R C E L

MARCELDEKKER,
INC.
D E K K E R

NEWYORK = BASEL HONGKONG

ISBN: 0-8247-0041-4
The publisher offers discounts on this book when orderedin bulk quantities.
For more information, write to Special Sales~rofessionalMarketing at the
address below.
This book is printed on acid-free paper.

Neither this book nor any part may be reproduced

or transmitted in anyform
or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage and retrieval system,
without permission in writing from the publisher.
W C E L DEKKER, INC.
270 Madison Avenue, New York, New York 10016

ht~:jjww~
dekker.corn
Current printing (last digit):
1 0 9 8 7 6 5 4 3 2 1

$TATE$ OF

CA

Modeling has become a well-developed discipline, and modeling research

on agricultural processes and operations is being conducted in many locations throughout the world. One of the most productive aspects of this work
is that it makes it possible for workers at distant locations to cooperate by
obtaining a copy of a program, and applying it
to their own agricultural
environment. The speed and inexpensiveness of the internet have also enhanced communication. This book is designed to further such interaction by
presenting the recent workof many of the top agricultural systems modeling
researchers in the world.
Models and simulation are mentioned throughout this book, so brief
definitions are necessary. We think of the model of a system as the set of
equations and rules that quantitatively describe the operation of the system
through time. Simulation is the process of solving these equations within
the rules with changing time, that is, simulating or mimicking the performance of the system over time by calculating values ofthe variables at each
series of time steps.
The process of modeling and simulation in agriculture is unique. First,
most agricultural problems require simulation
of continuous processes, as
well as discrete events. Second, agricultural problemsface the possibility of
discontinuitie~,such as those caused by a killing frost or a severe drought;
changes in the process function, such as shifting from daytime to dark reiii

iv

reface

spiration; and phenology changes, such as the plants

switch from leaf
expansion to seed production.
In addition, agricultural models usually seek to mimic real-life situations, so the effect of economics
is often a partof the problem. High-capacity
fieldequipmenthelps to serviceagronomic crops in a timely manner, to
avoidyield losses due to late operations. However, the fixed cost of the
equipment might be more than the savings from yield losses. Models that
aid in selection and design of field equipment for crop agriculture are presented in chapters on selection of machinery (Chapter 14) and on simulation
of a whole-farm system of specific equipment, crops, and operations that
must be carried out in a timely fashion (Chapter 15). As irrigation water
becomes more and more expensive, decisions are necessary, not only about
timing of irrigation scheduling, but also about the amount of irrigation water
to apply. Chapter 9 addresses these questions.The simulation of evaporation
of water from the soil and from the leaves of plants has been the subject of
a great deal of research, and a thorough review is presented in Chapter 7.
The amazing price/performance facts of todays personal computers,
as well as the continuing decrease in this ratio, make the equipment cost
factor for modelingatrivialquestion.Today,equipmentcostsareminor
compared to software development costs. Software engineers are helping to
ease this cost problem, by developing various new
software tools. Objectoriented programming systems(OOPS) are nowfavored for beginning a new
simulation project. Such systems are generally thought of
as a formalized
extension of theconceptofmodularprogramming,throughthe
use ofa
series of subprograms and a relatively brief mainor executive program.This
method is covered in Chapter 17.
ost of the models described here are one-dimensional in the sense
that time is the main variable. However, in many situations, the
space or
areal dimension is also important, as in the growth of the amount of space
occupied by an insect, or changes in the energy use over time, and in particular areas. This latter example is covered in Chapter 4, in which datafrom
India are presented.
any models of agricultural systems programmed by graduate students
have been subsequently laid to rest on the thesis shelf when students graduated and went on to new pursuits. A smaller number of models have been
writtenandrewritten,andarecontinuallybeingrevisedandreleasedin
newer versions, as with any commercial software. These are models that are
used by a number of researchers and teachers (including extension specialists), often in several countries throughout the world. This book encourages
the development of more useful models, to be shared with other scientists
and engineers working in agriculture. Examples of models that have been

reface

used by a substantial number of researchers are CROPGRO (Chapters2 and

18) and GRAZE (Chapter 10).
Many people in agriculture wonder why more of these models are not
used directly by growers on the farm; implying that if they are not used on.
the farm, then the cost of the development was not justified. However, this
overlooks the fact that historically most agricultural research involved finding out how a particular plant, animal,
or machine system behaves under
variedconditions,andmodels
were developed from thatresearch.When
questions came up about the possible
effect of climate changeon agricultural
crops, models of the major grain crops were already available. With rather
simple modi~cationsto account for the effects of carbon dioxide concentration, the models were the quickest and least expensive method
of estimating
the answers to these questions.
In addition to the previously mentioned reasons for the value of developing crop models even before they have been applied
to a particular
farm, the second answer is that some crop models are being used
on the
farm today. G O S S ~ M / C Othe
~ ,cotton growth and management model
described in Chapter 8, has been used by many farmers for about 10 years,
and the California program called CALEX, while not strictly a simulation
program, has alsobeenusedby
farmers andtheirtechnicaladvisersfor
several years. The growth of the profession of crop and pest management
consulting is an indication that the use
of crop modelsbythisgroupof
experts will probably be more widespread than the direct use of models by
farmers themselves. In addition, many of these models will be modified and
adapted for easier use with new software methods such as the various Visual toolsfordevelopingGraphicalUserInterfaceswithparticularprograms. Chapter 1 reviews important principles by which to adapt these researchmodels to farm use,according to what is importanttothe
farm
manager at the time. We predict growing use of agricultural models on the
farm with private or public extension workers involved in a great many of
these uses.
Other specialized methods of modeling are covered in this book, including the older linear programming method (Chapter
5), and the newer
neural network theory (Chapter16). Linear programming is a unique method
of finding specific solutions to problems that can be formulated to meet the
criteria of the technique (briefly, linear constraints, costs, and non-negativity). It is used routinely in many farm and agribusiness applications. Neural
network theory is much newer, and has been applied only in the past few
years. Linear programming works when the mathematical relationships are
linearandknown.Neuralnetworkmethodsworkwhenthere
is agreat
amount of data on inputs to and outputs from the system but the mathematical relationships are not known. Another newer technique-expert sys-

Preface

tems, or decision support systems-is presented here as a method of finetuning the parameter values of a crop model which includes the process
of moisture movement in the
soil (Chapter 6). Funda~entalmathematical
methods applied to biological processes, such as moisture and gas movement
throughthestomataofaleaf,arecovered(Chapter
3 ) to roundoutthe
mathematical and computational methods.
Modeling of livestocksystems probably does not have as long a history
as that of crop systems, because modeling livestock
on pasture involves
modeling the pasture as well. Four very different approaches are presented.
First, a fundamental method of modeling, using differential equations and
classical solutions for those equations, is well documented in Chapter 11.
Then, a dairy cow is simulated with the usual time steps and physiological
processes, but the feed supply does not include growth modeling of a crop
to grazing (Chapter
Chapter 13 shows how the detailed process of the
animal chewing on and consuming theforage can bemodeled and simulated
as part of the animal-pasture interaction. Chapter 10 uses the physiological
or process approach, with numerical integration, and models the growth of
the pasture and harvesting of the pasture by the animal (or
stopping this
grazing for a time) as interacting processes.
As the population of the world continues to grow, and with current
grainsurplusesbecomingalmostnonexistent,agriculturalproduction
has
become very important. Modeling and simulation of
food production systems will be increasingly used in the future, to help meet the global need
for food. We hope this bookwill help to spread the knowledge of agricultural
systems modeling and simulation to meet this task.

Robert M. Peart
R. Bruce Curry

Preface
Contributors

iii
ix

1. Humanization of Decision Support Using Information

from Simulations
John R. Barrett and Mark A, Nearing

2. Simulation of Biological Processes

James W Jones and Joep C. Luyten
3.

Using Mathematics as a Problem-Solving Tool

J. Robert Cooke
Integrating Spatial and Temporal Models:
An Energy Example
B. S. Panesar

5. Modeling Processes and Operations with Linear

Programming

113

Robert S. Sowell and R. C. Ward

Expert Systems for Self-Adjusting Process Simulation

157

Steven J. Thomson
vii

viii

7. Evaporation Models
John L. Monteith
8. Simulation in Crop Management:
G O S S ~ M / C O ~
Harry E: Hodges, Frank D. FVhisler, Susan M. Bridges,
K. Raja Reddy, and James M. Mc~inion
9.

Integrated Methods and Models for Deficit Irrigation

Planning

235

283

Neil L. Lecler
10. G W E : A Beef-Forage Model of Selective Grazing
Otto J. Loewer, Jr.

11. Dynamical Systems Models and Their Application to

Optimizing Grazing Management
Simon J. R. Woodward

301

419

12, Modeling and Simulation in Applied Livestock

Production Science

Jan Tind Sgrenson

13. ThePlant/Animal Interface inModels of GrazingSystems495
M. lierrero, Je B. Dent, and R. H. Fawcett
14. Field Machinery Selection Using Simulation and
Optimization

543

John C. Siemens
hole-Farm Simulation of Field Operations:
Object-Oriented Approach
Harbans Lal
16. Funda~entalsof Neural Networks

567

597

~ i l l i D.
a ~Batchelor
17. Object-Oriented Programming for Decision Systems

629

John Bolte
18. Simulation of Crop Growth: CROPGRO Model
Kenneth J. Boote, James W Jones,
and Cerrit Hoogenboom

651

Index

693

arrett Agricultural and Biological Engineering Department, Agricultural Research Service, U.S. Department of Agriculture, Purdue University, West Lafayette, Indiana.
atchelor AgriculturalandBiosystemsEngineering,Iowa
ioresource Engineering Department, OregonState University,
Corvallis, Oregon

oote AgronomyDepartment,UniversityofFlorida,Gainesville, Florida

ges Department of ComputerScience,Mississippi

University, Mississippi State, Mississippi

State

rt Cooke Department of Agricultural and Biological Engineering,

University, Ithaca, New York

e cur^ AgriculturalandBiologicalEngineeringDepartment,
University of Florida, Gainesville, Florida
ent Institute of Ecology and Resource Management, University
Edinburgh, Edinburgh, Scotland

of
ix

Fawcett Institute of Ecology and Resource Management, University

of Edinburgh, Edinburgh, Scotland
Institute of Ecology and Resource Management, University

, Edinburgh, Scotland
arry F. Hodges Department of Plant and Soil Sciences, Mississippi State
niversity, Mississippi State, Mississippi
GerritHoogenboom
Department of BiologicalandAgriculturalEngineering, University of Georgia, Griffin, Georgia
James W. Jones AgriculturalandBiologicalEngineeringDepartment,
University of Florida, Gainesville, Florida
arbansLal

Pacer Infotec Inc., Portland, Oregon

Neil L. Leeler Department of Agricultural Engineering, University of Natal, Pietermarit~burg,South Africa

Otto J. Loewer, Jr.

College of Engineering, University of Arkansas, Fay-

etteville, Arkansas

Joep C . Luyten AgriculturalandBiologicalEngineeringDepartment,

university of Florida, Gainesville, Florida

. McKinion
search
Service,
Mississippi

CropSimulationResearchUnit,AgriculturalReU S . Department of Agriculture,

Mississippi
State,

ohn L. Monteith Institute of TerrestrialEcology,Penicuik,Midlothian,

Scotland

ark A. Nearing AgriculturalandBiologicalEngineeringDepartment,

gricultural Research Service,U.S. Department of Agriculture, Purdue University, West Lafayette, Indiana
S. Panesar School of Energy Studies for Agriculture, Punjab Agricultural University, Ludhiana, Punjab, India

obert M, Peart AgriculturalandBiologicalEngineeringDepartment,

niversity of Florida, Gainesville, Florida
Raja Reddy Department of Plant and Soil Sciences, Mississippi State
University, Mississippi State, Mississippi
John C , Siemens AgriculturalEngineeringDepartment,University
Illinois at Urbana-Champaign, Urbana, Illinois
JanTindSfirensen
Department of AnimalHealthandWelfare,Danish
Institute of Animal Science, Tjele, Denmark

of

Contributors

xi

Robert S. Sowell AgriculturalandBiologicalEngineeringDepartment,

North Carolina State University, Raleigh, North Carolina
Steven J. Thomson U.S.CottonGinningLaboratory,AgriculturalResearch Service, U.S. Department of Agriculture, Stoneville, Mississippi

.C. Ward

Colorado State University, Fort Collins, Colorado

ler Department of Plantand Soil Sciences,Mississippi

Mississippi State, Mississippi
oodward

AgResearch
Ruakura,
Hamilton,

New Zealand

This Page Intentionally Left Blank

arrett and Mark A. Nearing

Agricultural Research Service, U. S. Department of Agriculture,

Purdue University, West Lafayette, Indiana

Agricultural researchers have tried from the turn of the 20th century
to develop data and information bases that when interpreted can be
used to improve the management of agricultural systems. Early on,
researchers recorded and published descriptions that, for the most
part, were observations of situations involving abnormal and unusual
occurrences. With the invention of adding machines and calculators,
scientists began to define and determine relationships. With the advent of computers, engineers began to consider continuous and discrete happenings with respect to time. Beginning in the late 1950s,
descriptive and mathematical modeling of processes evolved. These
mimics were called simulations. Figure 1 is suggested, somewhat
with tongue in cheek.
In the rnid-l960s, the concept of systems dynamics emerged,
facilitating time-related representations of process flows and interactions. During the
systems dynamics became formalized. Rel

Barrett and Nearing

FIGURE1 Simulating is stimulating!

~nementscontinued through the 1980s, primarily in computer programming techniques, in verification, validation, and evaluation,
Simulation of dynamic agricultural and industrial systems has become an integral part of agricultural science. Increasedunderstanding
of ecosystem interactions,as influenced by the environment and management practices, has greatly expanded the potential for decision
support systems. In the mid-1990s we experienced the emergence of
the era of information technologies.
Bridging the gap between simulations that have been developed
to mimic and describe agricultural processes and procedures and the
use of these simulations to supply informa~onthat supports decisions being made by managers is difficult and challenging. Supplying
information to managers involves interfacing between computer output and the people who are the information users (see Fig. 2).

II.

A first point is that all decisions are made in a person's mind. Some
human makes each, any.and all decisions, either directly or indirectly.
Further, decisions are always based on whether to continue doing
things as they are being done or to change to a different procedure.
And, people who control the resources have the authority to make
the decisions. The icon of Fig. 3 emphasizes all this.

Humanization of Decision Support

Knowledge
Engineering

Domain
Relationships

FIGURE2 Information system and decision making.

FIGURE3 Where decisions are made.

ing Process and

The mind, i n c l u ~ n gthe brain and sensory mechanisms for perception, functions to process and store symbolic expressions. An individual feels, perceives, thinks, remembers,and reasons in anadaptive
conscious and nonconscious manner. One way to think of this is that
the conscious mind is the interface between the source of symbols
that have been sensed and no~consciousmemory where they are
stored. Contained in a persons memory is everything that has been
observed and learned from all experiences, right or wrong, good or
bad. Certainly some things have become buried so deeply that they
are obscure. Stored in memory may be data, information, and/or
knowledge. Words are stored symbolically.
In resolving a problem, the contents of memory are reviewed,
perhaps intuitively and ordered according to importance, and alternatives are identified and automatically ranked. Risks are considered.
Short- and long-term consequences are explored, and choices are
made. Several matters may be thought through interactively and
simultaneousl~
When an individual is confronted with a problem to be solved,
several things happen. Some considerations occur almost instantly
and some are more protracted. What cognitive actions occur, the order inwhich they occur, and whether some of these actions recur are
a function of experience and knowledge of the domain where a decision is to be made. Most important, they are not a function of how
others see the problem, but of how the individual with the problem
sees it.
Mostly decisions are choices to continue a course of action orto
take a divergent course. Many preparatory dilemmas are resolved
over time before a decision is made to take action. Tpically when
an individual is confronted with a new situation or one where he or
she wants the action taken to be more effective, an attempt will be
made to review the situation. Thus, a fresh look may be taken to
reevaluate and reduce risks.
The blackboarding process that has been developed for computer information searching and problem solving is a mimic of what
in complex fashion goes on in a humans mind. That is, several problem areas may be addressed simultaneousl~The processing goes on
night and dayasleep or awake, consciouslyand nonconsciously until
the equations are complete or the situation is resolved to the point
that risk is adequately reduced and it is felt that action can be safely
taken. Then the resolution of the dilemma becomes evident. This may

umani~ationof Decision S u ~ ~ o ~

Benefits

FIGURE4 Decisionmaker.

occur in the middle of the night as if out of a dream or at some other

time.
All of this is important to an understandingof how simulations
are potentially excellent sources of important information that can
support decision making. Symbolically,
the interaction of high impact
areas is portrayed in Fig. 4.

So what does all this mean to decision support? Foremost, the job is
easier than may have been thought. Recognizing that decisions are
always made in a human's mind, support of the decision-making
process can.best be accomplished by supplying missing or incomplete
data or correcting inaccurate information about interactive critical areas in logic and achieving wording that is understandable and acceptable to the decision maker. This should point out the methodology that potentially can be used in any decision support process.
Simple? Not necessarily, because we are instilled through training andexperience to view problems from scientific perspectives,not
"from the applications-oriented viewpoints of people who manage
businesses. Scientists use the scientific method, which gives verified

and

Barrett

Nearing

and validated information with defined error, while others seek the
interpretation of information and data in changing situations. Scientists think in terms of probability, and decision-making managers look
to possibility.
A.

Example Sources ofInformation

Excellent farmers who produce grain said that the following sources
are most used for needed information (Barrett and Jacobson, 1995):
Networking with other experts.
Suppliers of chemicals, equipment, seed, etc.
Extension and university specialists, trusted, with no profit ties.
Consultants, although rarely used.
Manufacturers technical departments.
Publications, trade and other.
Specification statements and sheets.
Computers for grain price and weather information.
Note that this list does not include social contacts.
B.

ManagerialTime

These same excellent grain farmers explained managerial time by

pointing out that they are more receptive to information during certain periods of the year. Assume that a season begins with the end
of harvest, targeted in the Midwest to be about the first of November.
Following this are
1. The postharvest period, which lasts about 2 months, or until
the first of the new calendar year. Postharvest is when decision support can best occur, as during this time all plans
are finalized for the next season, considering prior experiences, especially the immediate past growing season.
2. Preplanting or the
month period used for getting everything ready for planting.
3. Planting, targeted to be 2+ weeks.
4. The growing period, which lasts about
months.
5. Harvest, which occupies the remaining l + month.
To support decisions with maximal effect,i~ormationshould be presented at times when information users are most receptive. In the
case of grain farmers, this would be from late harvest until the end
of the year for strategic planning, and/or as they walk out the door

umanization
Supportof Decision

to accomplish an operation such as planting, tilling, or irrigating. This

latter is tactical and short-run operational.
G.

Computer Programs

Computer programs including spreadsheets, expert systems, decision

support systems, simulations, and similar items can be excellent
sources of information for decisionsupport. For more information on
knowledge engineering in agriculture, see Barrett and Jones (1989).
Unfortunately, there has been lirnited acceptance of computer programs by information users. Why? Perhaps it is partly because the
information as presented is not what is needed. Of course there are
other reasons. Computer output must be in wording and logic that
can be understood by the information users.
The founders of artificialintelligence-Minsky,Feigenbaum,
Englemore, and others-stated emphatically, in personal communications about 1983, that energies should be spent in defining problems, that representative clients should be included on research
teams, and that programming, language development, information
storage and retrieval, machine learning, and other mechanistic computer-related developments should be left to specialists. Thegreatest
challenge would be to develop software programs that would actually be used. Advice was to specifically:
1. ~a~ needs a s s e s s ~ e n t s . Determine from discussions with
the proposed clients that there is a real and economically
viable need for program development and that there is a
chance of meeting that need. Many programs are produced
only because there is a body of knowledge that can be
expressed.
Involveusers. Both users of programs and users of the information generated should be included on developmental
teams from the beginning, not simply to follow progress and
to evaluate output.
3. Clearlydefine i ~ ~ o r ~ a n t ~ r o bProgram
l e ~ s . logic needs to be
expressed from the perspective of the people who have the
problem. Problem definition must be from an information
users viewpoint and not be limited by the domain
specialist.
4. Not worry about software and ~ r d w a r e . Developers were advised to leave software writing to specialists and consider
program needs, economics, and availability when selecting
hardware.

5. Addresstheneeds of ~otentialusers, Usersmay be farmers,

extension service specialists, scientists, agribusiness people,
or ctther clients. They seek information and interpretations.
To be understood, outputs should be in the language of the
information users, not that of the domain experts.
'
Allow for ~ a i n ~ e n a n ~ The
e . 80/20 rule characterizes the development process. It takes 20% of the effort to develop a
system or program, and the other 80% for validation, delivery, and maintenance.
7. Know the~enefactors. Keep in mind the people who are
funding the work.

Simulation is the process of using a model, or models, dynamically

to follow changes in a system over time. This is dynamic mimicry.
Models maybe descriptive and/ or mathematical. Technically, models
are equations and rules defining and describing a system, Simulation
involves calculating values over time in dynamic fashion. For thorough understanding of modeling and simulation, look to Peart and
Barrett (1979) and Barrett and Peart (1982).
Almost all phenomena related to agriculture have been simulated, some accurately. Everything-for example, biochemical
pathways,erosion, water runoff, crop drying, pesticideleaching, crop
growth and development, harvest, transport, residue decomposition,
the aspects of cashflow and economics,costeffectiveness,
weed
stress, pest (insect, weed, disease) management, ~ a c h i n selection
e~
and management, animal production, dairy herd improvement, hydrology, pollution, irrigation, biomass burning, global warming,
weather-has been simulated..Most of these simulations came out of
a scientist-researcher's desire to express his or her understanding of
system interactions. Nevertheless, simulations are resources that can
be used to supply data, information, and knowledge to decision makers in agriculture (Fig. 5). For example, a crop growth and development simulation can be used to predict the amount of dry matter in
a plant or crop at the time of harvest, which can berelated to residue
left on the surface after harvest. Considering dispersal and decomposition, the residue cover at planting can be predicted and related
to the potential for erosion or weed suppression. All of these relationships can be simulated to help a farm manager make strategic
and procedural decisions. The difficulty lies in expressing relationships of critical factorsin a way theinformation user can understand.

FIGURE5 Data, information, and knowledge are resources.

Research carried out to support decisions in grain production was

used as an example by Barrett and Jacobson (1995). Excellentfarmers
were observed over a 2 year cycle (rotation of corn and soybeans) to
determine the factors critical to their success (Rockart, 1979; Bullen
and Rockart, 1981). These factors were related to the best management practice areas. A computer process/program was developed
that emphasizes the most important management practices where optimization can occur. These areas are listed below in the experts priority to show where simulations can be used to supply information.
Cash pow. This involves timing and the dynamics of selling the
crop, buying resources, attaining maximum return in dollars
and yield, achieving sustainability in farming, making major
payments, taxes, interest, etc. From a producers viewpoint,
economics per se is a subset of cash flow.
Conditions are basic to all thinking and decisions.
Management must always be aware of conditions, local, regional, and international, including rain; the effecton scheduling; surface conditions; subsoil levels; if and when to irrigate. Managing with respect to surface,soil,
and grain
moisture is paramount to success.
Here equipment includes maintenance, manner
of operation, repair, replacement and change, purchase, labor
required for use, innovations.

Barrett and Nearing

~ n ~ o r ~ a t i o Information
n.
must be continuously searched for

with respect to prices, yield, policies, inte~ationalaffairs, tillage, variety performance, new technologies, etc. Networking
is useful here.
Cropeffects. Status of growth and development, local and regional; performance considering moisture and soilstatus;
vigor after emergence; stress factors; stability at harvest; shatter; size; stamina; etc.
Weedsuppression.
Existing and potential pressures related to
tillage; effectiveness of chemicals, spray s c h e d u ~ g etc.
, The
producers were observed to think of conventional tillage, reduced tillage, and no-till as variations of seedbed preparation
directly affecting potential weed pressures.
~ l a n t i n ~Varieties;
.
scheduled over
week period; when to
start with respect to harvest; tillage; labor; use of planter or
drill; etc.
~ e r ~ i l i t y Maintenance,
.
soil testing, scheduling, etc.
Harvest. Schedule over l + month period; preparation; storage;
labor; timing relative to planting schedule; etc.
Erosion. Environmental stewardship, sustainable agriculture,
sustainable production, conservation, compliance, etc.
Diseases,insects.
Scouting; chemical selection; spraying; etc.
O t h e re n t e ~ r i s e s .
All observed expert farmers maintain an
associated non-production-agriculture business enterprise
where resources are used interactively.
Labor, personal preferences, other crops, other factors.
Any information from simulations about interactions of the above
areas can potentially improve decision making.
r o ~ r a mto Supply l ~ f o r m a t i o ~

To continue the example being discussed, after considering the

management practice areas that have been identified,aprototype computer decision support system of programs was assembled
that includes an interface between the computer user and the physiologically based soybean and corn crop growth and development
simulatio~s, en~ronmentalquality simulations, hydrology simulations, and weed and herbicide effect models. There is also an additional program structure to compute profitability. For detailed discussion, see Jacobson et al. (1995).
The important problem to be solved is to complete the interface
between the cornputer as a source of data and information and the

Humanization
Supportof Decision

11

information user, who is usually not the keyboard user. Relating the
outputs of these programs with theinformational needs of a manager
of production is not a straightforward task. The information being
sought relative to the critical areas must be extracted and phrased in
words and logic appropriate to the decision maker.

Early erosion simulations were empirical equations. Later, processbased models were developed. Information on rates of erosion related
to land management on a specified site continues to be objectively
sought.
A.

Universal Soil loss Eqwation

The first widely used erosion simulation was the Universal Soil Loss
Equation (USLE) (Wischmeier and Smith, 1978). This statistically derived regression model was developed to estimate average annual
soil lossfrom a hillslope.USLE does not address sediment deposition
on the toe slope of the hill, the amount of sediment leaving a field,
or the composition of the sediment eroded.
USLE has been the basis for the development and implementation of conservation practices throughout the world. It has been
applied to nearly every acre of cropland in the United States. The
primary reason for its success is its simplicity. The mathematics can
be done on a calculator, on the back of an envelope, or in the head.
The answer is also simple, being a single average annual soil loss
value for a hillslope. USLE is easily implemented and easily integrated with large and complex simulations and GIS applications.
evised Universal Soil loss Equation

The Revised Universal Soil Loss

Equation (RUSLE) (Renard et al.,
1996) is a computer version of USLE. RUSLE includes additional process descriptions such as temporal changes in soil susceptibility to"
erosion and the time decay of surface crop residue.
Using RUSLE involves a computer, as it requires extensive data
input, including access to files for plants, climate, and management
systems. It makes simulations on a daily to biweekly time step for
various processes that influence erosion, such as residue decay, tillage, seasonal soil erodibility and plant biomass production. The results of RUSLE are allegedly better than those of USLE because timevariable factors are considered that USLE does not include and

because RUSLE was designed to better predict erosion on rangelands

with minimum tillage conditions. Yet overall, RUSLE provides the
same level of information as USLE, i.e., average annual soil loss on
a hillslope.
rosion ~ r e ~ i c t i oProject
n

The Water Erosion Prediction Project (WEPP) model (Nearing et al.,

1989; Flanagan and Nearing, 1995) is composed of a complex set of
process-based models. It provides information on soil loss, deposition,%
sediment leaving the field, and the characteristics of this sediment on a daily, monthly, annual, and average annual basis. This is
important in terms of off-site environmental impact. It provides information on soil water content over time and surface water runoff
rates and amounts. It provides information
the size distribution
of sediment that leaves the field. Characterization of sediment is important for quantdying chemical transport.
The WEPP model predicts crop yields. WEPP also gives spatial
and temporal distribution ~ f o r m a ~ o
onn erosion. This should be of
benefit in meeting the need$ of the precision farmer. On the other
hand, WEPP is data-intensive, both for input and for output. Interpretation of the WEPP output information requires an erosion expert.
It will be interesting over the next few years to observe how
these simulations are used. The challenge willbe to make the models
effective in supplying information that supports decisions.

Uses of $osion Sim~lation

Erosion simulations are used for conservation planning, engineering

design, erosion inventories, and government regulations. USLE, and
now RUSLE, are the worldwide standards for evaluating conservation practices, Farmers in the United States, with the aid of soil conservationists, use these in estimating soil losses from their fields as
they choose among strategic and tactical alternatives.
Erosion simulations are used extensively in the design of engineering structures such as terraces, diversion ditches, sediment catch
basins, reservoirs, and drop structure plunge pools. They are also
used as the basis for local, state, and national soil resource inventories, such as the USDA Natural Resource Inventory, which is conducted on a five-year basis.
The USLE was mentioned in the language of the 1985 USDA
Farm Bill to define highly erodible lands. These lands were then tied
to farm price support and loan program regulations. The 1985 farm

wmanizati~nof Decision S w ~ ~ o r t

13

bill was inst~mentalin both identifying, on the broadest scale, the

field areas with high erosion potential and e n c o u r a ~
compliance
~~
with conse~ationguidelines on erosion-s~sceptiblelands. Basically
fields that were identified as highly erodible had to be placed into
approved conservation plans in order for the producer to be eligible
for federal farm programs.

The effectiveness of more precise ( p r e s c ~ ~ ~farming

o n ) can be evaluated through the use of s i ~ u l a t i o n
and
~ economic analysis. Benefits

and

14

Barrett

Nearing

will vary among farms, with the technology used,and with the levels
of management abilities. The underlying ideas of precision farming
are solid and consistent with the concepts of good management. One
farmer of the grain production example said that he had been practicing prescription farming all his life. His only questions were: How
precise should he be, what size areas should he consider,and whether
and when he should change! He said that practicing precise, sitespecific farming was what made excellent management.
Weed Suppression Example

Weed control, or suppression, is complicated when viewed from a

grain producers perspective. To show how farmers think about weed
suppression, consider that weed control is a continuous and ongoing
problem area. The potential for excessive competition is constantly
being evaluated. Critical actions affect sustaining production and sustaining agriculture and environmental stewardship. Fall tillage, the
purchase of herbicides, preplanting tillage and herbicide application,
planting time tillage and herbicide application, postplanting tillage
or herbicide application,labor utilization, safety,equipment purchase,
repair and management, continuous and crisis control, etc., all interact with each other.
Weed identification, herbicide selection, and cost effectiveness
are rather straightforward. Their interaction is where simulations can
fit together to potentially supply critical information. Information
users want to know how the resolution of each dilemma over time
will affect yield and profitability and if they can sustain production
over their lifetime and for generations to come. Simply stated, the
question is If I use more or less, here or there, will it pay? where
pay does not necessarily refer to dollar amounts.

F e ~ i l i z a t i o nExample

Relative to prescription (precise) farming, farmers in the Midwest are

asking for help in determining whether changes in fertilization application rates, fertilizer types,or application methodologies will improve their profitability. If they are good producers, they monitor and
maintain fertility levels on an ongoing, long-term basis according to
characteristics of the fields-slopes,soiltypes,
and so on. Should
their methods be more refined? Simulations can supply the answers
to their site- and situation-speci~cquestions.

Humanization of Decision Support

C.

15

DairyManagement

In similar thinking, dairy production can be improved by using simulations to evaluate the effect of changes in herds, both numbers and
quality of animals. This combination of cow and herd genetics, resource management, economics including sales methods, government
programs, and quality control is basic to good business management.

The application of site- and situation-specific production concepts is

destined to bring a whole new type of management to agriculture,
one based on data and information technologies. To effectively implement more precise (prescription) production, the variability associated with small units (areas, cows)must be determined so that production yields and profitabilitycan be predicted along with the
potential for beneficialresponses to input resources. Thishas the potential of extending into all facets of agriculture.
OCESS O F DECISION SUPPORT

From an information users viewpoint, there are several somewhat

simplistic considerations that will help improve decision support. After the computer has been introduced as a source of help (not trivial),
it becomesimportant to establish a believable base, A simulated base
must be shown to replicate reality-a situation that the information
user knows to be true. Then it becomes a matter of showing the
results of each proposed change. If a ranked group of alternatives are
presented as program output, each alternative will be viewed as information that when added to all the other information in memory
supports change from the procedure being presently used.
Acceptance canbe improved by leading information users to see
that their personal data have been properly and accurately input,
including weather, and that the yield, profit, and corresponding results predicted are what they observed to have happened. Then the
person who is making the decision willbe ready to accept as accurate
the effects of changed inputs. Simulations come into play as they
mimic procedures and show the potential results of changing resource use.

It is a great challenge to develop simulations and use existing ones

objectively to produce information that supports decisions. Other-

wise, in most cases itis still necessary to determine what information,

known and unknown, is critical to the resolution of problems and the
improvement of decision making management. It is necessary to determine what available simulations are relevant and how to get the
required information out of them. It is necessary to interface between
simulations and other software, to interface output with information
users, to determine where data and information bases can be found,
to determine what can be easily input inuser-specific manner,to limit
the items that must be leftto theinformation user to supply, and how
to play believable what-if games.

Muchpolitical,social, and scientific ,emphasis is being placed on

global warming, environmental pollution, chemical leaching, pesticide and herbicide runoff erosion, conservation, and other qualityrelated matters. Correspondingly, we now have the ability and the
opportunity to bring the processes of decision support together with
simulation and the emerging methods of information technologies to
better bridge the gap between research and practical applications. It
is exciting to catch glimpses of the future and see that a farmer's
sustainable production can be balanced with the env~onmentalist's
sustainable agriculture so that we can be better environmental stewards while enjoying a high quality of life.
The future is unlimited! Grab the brass ring!

Barrett, J. R., and B. M. Jacobson. 1995. Humanization of decision support

for managing U.S. grain (soybean and corn) production. In:Proc. 2nd
IFAC/IFIP/Eur~gEngWorkshop on AI in Agriculture. Wagenin~en,The
Netherlands, 29-31 May, pp. 3-13.
Barrett, J. R., and D. D. Jones. 1989. ~ n o ~ Z e ~ g e ~ n ~ n in
e eA~icuZture.
ring
''Monograph No. 8. St. Joseph, MI: ASAE.
Barrett, J. R., and R. M. Peart. 1982. Systems simulation in US. agriculture.
In: Progress in ~ o ~ e Z Z i nSimuZation.
g
F. E. Collier (Ed.), London: Academic
Press, pp. 39-59.
Bullen,C. V., and J. E Rockart. 1981. A Primer on Critical Success Factors.
Center for Information Research No. 69. Sloan Schoolof Management WP
No. 1220-81, Cambridge, MA: Massachusetts Institute of Technology.
Flanagan, D. C., and M. A. Nearing. 1995. USDA-Water Erosion Prediction
Project: Hillslope Profile and Watershed Model Documentation. NSERL

Report No. 10. West Lafayette, IN: USDA-ARS, National Soil Erosion Research Laboratory.
Jacobson, B. M., J. W. Jones, and S. M. Welch. 1995. Decisionsupport system
to assess agronomic, economic,
and environmental impacts of soybean and
corn management. ASAE Paper No. 95-2696. St. Joseph, MI: ASAE.
Nearing, M. A., G. R. Foster, L. J. Lane, and S. C. Finkner. 1989. A processbased soil erosionmodel for USDA-Water Erosion Prediction Project technology. Trans, ASAE 32:1587--1593.
Peart, R. M., and J. R. Barrett. 1979. The role of simulation. In: ~ o d ~ c a t i o n
of the Aerial E n v i ~ o n ~ ofe nPlants.
~
Monograph No. 2. St. Joseph,MI: ASAE,
pp. 467-480.
Renard, K. G., G. R. Foster, G. A. Weesies, D. K. McCool, and D. C. Yoder.
1996. P r e ~ i c ~ i nSoil
g Erosion by Water: A Guide to Conservation ~ l a n n i n gwith
the Revised ~niversalSoil Loss E ~ ~ a t i o n ( R ~ SUSDA
L E ) . Handbook No. 703.
Washin~on,DC: Government Printing Office.
Rockart, J. R. 1979. Chief executives define their own data needs. ~ a r v a ~ ~
~ u s i n e s Review.
s
March/ April, p. 81.
Wischmeier, W. H., and D. D. Smith. 1978. Predicting ~ i n ~ aErosion
ll
Losses:
A Guide to Conservation Planning. USDA Handbook No. 537. W a s h ~ ~ o n ,
DC: Government Printing Office.

This Page Intentionally Left Blank

Jones and Joep C. Luyten

University of Florida, Gainesville, Florida

Biological systems are made up of interacting chemical and physical

processes. Living systems are composed of many subsystems and
components, eachhaving its own unique characteristics and behavior
while contributing to the overall form and function of an entire system. These systems are highly complex; many components interact
sim~ltaneousl~
and their interactions are highly nonlinear or chaotic
in nature. These interactions and nonlinearities must be taken into
account when attempts are made to understand or predict system
behavior. Our understanding of these interactions is incomplete and
often guided only by empirical evidence of overall system behavior
instead of empirical data on processes that lead to overall system
behavior. Because of these complexities, the classical mathematical
methods used to study nonliving physical or chemical systems have
been inadequate for living systems.
Simulation based on quantitative models of biological processes
and their interactions can provide considerable insight into the behavior of living systems and into ways of managing these systems to
achieve specific goals.During the last several decades, computer sim-

ulation has proved to be a powerful tool in basic and applied biological sciences. The incredible growth and acceptance of personal computers during this same time period has made it possiblefor
simulation to become an integral part of biological researchin many
basic and applied fields. The purpose of this chapter is to present an
introduction to the concepts and techniques used in the simulation
of agricultural and biological systems with a few examplesthat demonstrate the approach.
Because of our own interest and experience in the simulation of
agricultural production systems (particularly crop systems), we focus
on the simulation of crop systems in this chapter and include examples related to plants, soil,and insects. Additional material on crop
s~mulationcan be found ina number of references, including Thornley and Johnson (1990), Leffelaar (1993),
Penning de Vries et al.(1989),
and Goudriaan and van Laar (1994). The conceptsand me tho do lo^
could have been presented equally well with examples on animal,
disease, or other biological systems (e.g.,see Curry and Feldman,
1987; Keen and Spain, 1990; France and Thornley, 1984; Odum, 1973,
and Dent and Blackie, 1979).

A system is a collection of components and their interrelations~ps

that have been grouped together for the purpose of studying some
part of the real world. The selection of the components to include in
a system depends on the objectives of the study and actually represents our simplified view of reality. Systems can be described as collections of mutually interacting objects that are affected by outside
forces (Pritsker,1995; Rabbinge etal., 1994). For example,
typical crop
models define the crop and soil rootzone as components that interact
in complex ways and are also affected by weather conditions and
management practices. Other models may define a leaf or a cell as
the system for which models are to be developed.
One of the complicating features of biological systems is that
they are hierarchically organized and can be studied at a number of
levels. Figure 1 shows hierarchical levels for crop systems. At different ~erarchicallevels, models and simulation analysis are being performed by scientists, engineers,and economists. Notethe parallel between research and models at each of these levels. The fundamental
goals and objectives of a model should guide one to determine which
hierarchical level to use.

t
N
G
E

E
N
T
I

FIGURE1 Hierarchy of models used in agriculture.

To consider the scope of a system, one must prescribe its boundaries

and its contents. The environment of system includes everything
except the components of the system. A system boundary is an abstraction of the limits of the system components, separating them
from the environment. The boundary may be physical; however, it is
better to think of it in terms of cause and effect. The environment
may affect the system in a number of ways, but the system does not
affect the environment. For example, there may be a flow of mass
and energy from the environment to the system that affects the systems behavior. Themodel of the system must take into account how
these flows affect the changes in system components. In contrast, the
environment is not affected by the system, and theenvironment does
not have to be modeled. For example, in a crop model, the crop is
affected by the temperature and radiation of the environment but
does not itself influence these environmental conditions. Of course,
crop processes do affect the microclimate and, in a very small way,
other meteorological conditions,but these are usually assumed to be
negligible for a number of purposes.

22

Jonesand ~ ~ y t e n

A model is defined as a mathematical representation of a system, and

modeling is the process of developing that representation. Conceptually the process of developing a model of a system is different from
and a prerequisite to computer simulation.

~ o r n ~ u t Sirnulation
er

Computer simulation includes the processes necessary foroperationalizing or solving a model to mimic real system behavior. Developing
computer logic and flow diagrams, writing the computer code, and
implementing the code on a computer to produce desired outputs are
necessary tasks in the simulation process. In practice, modeling and
computer simulation are usually interrelated. One should have in
mind techniques for i m ~ l e m e n t i nthe
~ mathematical model as it is
being conceptualized and developed. Some authors define simulation
to include the modeling process (Pritsker,1995).In this chapter, "simulation of biological systems" refers to the overall procedures outlined below, including mathematical model development and computer simulation to produce numerical model results for analysis.

Inputs to thesystem are factors in theenvironment that influence the

behavior of the system but are not influenced by the system. Inputs
are also referred to as exogenous variables,driving variables, or forcing functions. Inputs may vary with time, and numerical values for
inputs are needed for all values of time for which the model is to be
simulated by prescribed values, tables, or equations. Rainfall, temperature, and light are examples of inputs to crop systems. These
environmental variables affect the soil water and crop growth dynamics in a field but are not themselves affectedby the crop system. The
choice of components and inputs for various models may differ depending on model objectives and the availability of data. For example, if one chooses to continuously monitor soil water conditions, then
soil water could be an input to a crop model having no soil component. However, soil components are usually added because of the
difficulty in monitoring soil water at any site where one wishes to
apply the model. This is in contrast to the relative ease with which
rainfall can be monitored and used as input to a crop model with a
soil component.

iological Processes

23

Outputs from the system represent the characteristic behavior of

the system that is of interest to the modeler. For example, outputs
from a crop system model would include biomass of the crop and
water content of the soil component.
arameters and on st ants

Parameters and constants are characteristics of the components of the

model that are constant throughout simulated time. Usually constants are quantities with reliable and accurate values that remain the
same when experimental conditions are varied or when the model is
applied to different genotypes or organisms (France and Thornley
1984). Examples of constants are the molecular weight of glucose, the
gravitational constant, and the number of seconds in a day. Parameters, on the other hand, are quantities whose values are less certain
but are assumed to be the same throughout a simulation. For example, parameters may definethe functional response of photosynthesis
to light, the resistance to soil water flow,, respiration loss for tissue
synthesis, and the number of degree-days from crop emergence to
flowering.

State Varia~les

State variables are quantities that describe the conditions of system

components. These state variables change with time in dynamic models as system components interact with each other and with the environment. For example, soilwater content and crop biomass are two
state variables that change with time in most crop models. A model
may have one state variable or it may have many to describe the
various characteristics of a system that change with time. State variables are of critical importance because these are the dynamic characteristics of the crop or other system that are of interest to researchers.

The interrelationships between components in a system, and therefore between state variables in the system, exist because of various
processes. We sometimes use the term "process-oriented, to describe
models that describe the flow and storage of mass, energy, or other
substances. For example, the crop biomass state variable changes as
a result of photosynthesis and respiration processes, and the soil water state variable changes as a result of rainfall, runoff, percolation,

and evapotranspiration processes. A crop model is the set of mathematical relations~psthat describe the changes in state variables as a
result of the various processes that occur.
C o ~ ~ i ~ u omodels
us
are characterized by state variables that can
change smoothly over small time intervals and are not restricted to
integer values. Discrete models, on the other hand, are those in which
the variables describing the system states take on integer values. Crop
models are usually classified as continuous models. An example of a
discrete biological model is one that predicts the birth and death of
individual insects. The number of insects alive at any point in time
is a whole number. Discrete models usually require information on
the times required to complete activities, such as the time required
for an insect egg to hatch. In contrast, continuous models require
~fo~matio
onn processes such as the flow rate of material or energy
between components and between components and theenvironment.
Continuous models are usually represented by a set of differential or difference equations derived from the structure of the system
and the interrelationships among components. Some systems can be
modeled as either continuous or discrete, depending on the purpose
of the model. For example, the fate ~ i r t hdeath)
,
of individual insects
may be viewed as a series of discrete events. The population number
at any time would be a count of the total number of individuals, each
considered separately. Obviously, if the population is large, a lot of
computer time is required to keep up with each insect. One may also
view the insect population density ( n u m ~ eof
r insects per unit area)
as a continuous state variable for large numbers and when population
rocesses (births, deaths) occur smoothly overtime, and thus use
differential equations to model their dynamics.

Veri~cationinvolves the evaluation of the accuracy with which the

computer code represents the mathematical model and the programmers intentions. Verification also involves the careful checking of
mathematical manipulations, units and their conversions, and programming logic and code to ensure that neither the mathe~atical
model nor its translation into one or more computer programs has
errors in it.

Calibration consists of making adjustments to model parameters to

give the best fit between simulated results and results obtained from

measurements on the real system. Inother words, calibration involves

adjusting certain model parameters by systematically comparing simulated results with observations of state variables. Model structure
remains the same, and parameters are adjusted to more closely describe observed behavior, For example, suppose that partitioning of
new crop growth to leaves varies with variety all else being equal.
An experiment could be conducted in which the total crop and leaf
dry matter are measured and a leaf partitioning parameter estimated
for each variety by fitting simulated results to match observed data.
Calibration should be conducted only within the confines of a given
data set. In many cases, calibration is the only practical way to estimate some parameter values that are used in biological models.

Validation is the process of comparing simulated results to real system data not previously used in any calibration or parameter estimation process.The purpose of validation is to determine if the
model is sufficiently accurate for its application as defined by objectives of the simulation study. Simulated state variables are compared
with measured values of state variables. Usually in crop simulation
studies only a few state variables out of many possibilities are measured, and thus a complete comparison is usually not possible. Validation involves subjective judgment. First, the areas for comparison
must be selected. Then a measure of "accuracy" or "closeness of fit"
must be established, such as the final crop yield. The choiceof criteria
and important state variables for comparison should be based on
model objectives. Validation effortsare essential in the application of
crop models.

.
Simulation involves the development of a model and its use in gaining insight about the system itself or its management. The Simulation
approach usually entails the use of a number of steps, which are
summarized below. Following this summary, techniques for biological model development and implementation are presented along with
an example of a simplified crop model.

A clear statement of the reasons forundertaking the simulation study

is essential. All remaining steps in the simulation study depend on

this initial step. The problem to be addressed and the information to

be derived from the model should be explicitly identified. Although
this is one of the most critical steps in a simulation project, it is often
overlooked or is not clearly communicated. As a part of the statement
of objectives, there should be a clear definition of the intended end
product and the intended users of the models that will be developed
within a project. Without such documentation, project goals may remain poorly defined and reduce the effectiveness of contributors to
the project. This step is not as easy as it may seem. The objective of
the simulation effort should be used to determine the level of detail
needed in the model, the type of experiment that is needed, the type
of data that should be collected, and every other step that follows,
Broadly speaking, there are two fundamental objectives for biological simulation models. First, scientistsmay wish to obtain a better understanding of the behavior of a system, the interactions that
occur, and cause-and-effect relationships in a system. Thus, simulation of biological systems can help evaluate ones understanding by
testing hypotheses about its behavior, using models and well-defined
experiments (Boote et al., 1996; Sinclair, 1990; Monteith, 1990; Goudriaan and Monteith, 1990; Goudriaan, 1988). For example, a scientist
may be interested in developing a quantitative model of crop canopy
photosynthesis under varying temperature and carbon dioxide conditions to gain insight into possible crop responses to changing climatic conditions. A model would serve as the scientists hypothesis
concerning crop process and overall response, and experiments
would be designed specifically for determining the adequacy or inadequacy of the hypothesis. The end result of such a study could be
a model with improved capabilities for prediction,or it could lead to
a new research when a model and current hypotheses are shown to
be inadequate. This fundamental objective for modeling could be referred to as a scientific goal since the major expected result is an
increase in knowledge.
The second fundamental objective is more problem-oriented or
applied and could be referred to as an engineering goal. In this case,
the major goal relates to better prediction of system behavior for use
in improving control or management of a system. For example, a
research team may wish to develop a model for use in a computerized
i ~ g a t i o nsystem controller or to be used by water resource policy
board members to help them decide how to allocate water during
periods of drought. The end result of this type of project is a software
product or combination hardware and software product designed for
a specific application.

i ~ ~ o g i cProcesses
a~

27

It can be argued that scientific models can also be applied and

that there is really no distinction between these two goals. However,
for problem-oriented or engineering objectives, it is highly important
that the model predict the system behavior with an acceptable level
of accuracy. Thetype of model is not important as long as it is reliable
and performs as needed. In contrast, scientific models may not predict the systems behavior well at all, but the study itself could be
very valuable. It could demonstrate where knowledge is inadequate
and where additional research is needed. In this case, the form of the
model is important; it should be structured to represent current scientific understanding of the processes and used to help design experiments and data collection methods for testing the understanding.
Bothscientific and engineering objectives are relevant. General arguments about which is the most appropriate approach serve no useful purpose. Applied models of biological systems evolve as understanding is improved; scientific modeling is a highly effective method
for helping researchers improve their understanding of biological
systems.
One canthink of biological simulation efforts as being productoriented. The end product may be increased knowledge and an improved ~nderstandingof crop behavior, or the end product may be
a tool that is designed for application to specific problems. This characterization of objectives sets the stage for all activities a simulation
project.
efinition of the Syste

The components to be included in the study andthe system boundary

should be identified. Inputs and outputsshould be described. As additional information is gathered, it may later be necessary to redefine
the system. Or, for example, lackof available data may force a change
in the system definition.
C.

~ i t e r a t ~ rReview
e
and

Prior to model development, a review of information availability is

needed in order to evaluate the possibility of meeting objectives as
stated. In this step, the essential features of the system are established
and a conceptual model that will depend on the data available and
the ability to quantify processes in the model is developed.

In the model development step, diagrams may be used to represent

the components in the system and to summarize their interrelati~n-

ships. The mathematical representation of the system should be developed, including specific functions and relationships to be used in
the model. Experiments may be
needed to develop functions and
estimate parameters for the model. Themodel is translated into computer code for simulating the behavior of the real system. Computer
flowcharts are usually helpful in translating the model to computer
code and documenting the relationships between the model and the
code. Techniques formodel development are presented in Section IV.

The question of whether a model is accurate must be followed by

for what objectives? and as compared to what? In the final analysis, one must ascertain whether the cost of obtaining an additional
unit of accuracy is greater than the benefits to be gained in terms of
the study objectives. Accuracy is usually defined in terms of verification, calibration, and validation.
Dent and Blackie (1979) provide an excellent perspective on
model evaluation:
When we are confident that the behavior of the model is satisfactory, the formal validation process is over. Although a
formal process of validation is always recommended, the process of gaining confidence in the model is generally a slowly
emerging one over the period of model construction, through
formal validation, to application of the model. During this
time, assessment and modification will proceed, essentially
two sets of judgments being made on the way:
a. That the model is not different from the real existing system to a degree that will detract from the value of the
model for the purposes for which it was designed;
b. That if the model is accepted as being adequate then the
decisions made with itsassistance will not be m e a s u ~ a b l ~
less correct than those made without the benefit of the
model. Such an exactassessment is extremelydifficult
and cannot be other than subjective in nature. Subjective
judgments in this regard are, however, not without value
and this aspect of validation mayoftenbe
the most
relevant.
It is because of the difficulties associated with these two sets
of judgments that validation remains an elusive issue in the
simulation procedure. Statistical tests are available to deter-

mine whether the model behavior is differentfromrealsystem behavior to some stated level of significance. Theprecision of such tests should not overshadow the conceptual
problems in applying them. There is no mechanical procedure
that permits the conditions for acceptablevalidation by a
gle (or even a series of) statistical comparison(s)of the modelperformance against some recorded or measured data from
an existing system which the model represents. Any statistical
tests that may be carried out and which are positive in favor
of the model add to the model-builder's confidence that his
creation is functioning well.Theacceptablelevel
of confidence is achieved normally by a series of assessments and
subsequent modifications until such time as the model is to
be applied in support of decision making.

Sensitivity analysis involves exploring the behavior of the model for

different values of parameters. This is done to determine how much
a change in the value of a parameter influences the important outputs
from the model. Techniques for performing sensitivity analysis are
presented in Section VI.

The model application may relate to management of a system. In this

case, the model may be viewed as "complete," and its users may not
have participated in its development. Documentation of the model
and the computer code is essential. The original objective may be
related to research and the use of systems modeling to assist in a
more thorough understanding of the system. In this case, the use of
sensitivity analysis in guiding research efforts may be the intended
application.
Simulation steps should not be viewed as a strict sequential process. Rather, in most systems modeling and simulation studies, these
steps are repeated in an interactive fashion as information is gained
and progress is made.

.
Continuous system modeling is a process-oriented approach for describing the behavior of a system. Processes fall into three categories:
transport (or flow), transformation, and storage (Smerage, 1977).

Jonesand L ~ ~ e n

These processesare described by two classes of variables: extensive or

through variables and i n t e ~ s i ~ore across variables. Extensive variables are characterized by flow-through quantities such as mass, volume, electric charge, force, and heat flows. Intensive variables are
measures of energy intensities or potentials across system components. Intensive variables represent the driving force forthe extensive
variables. Examples are pressure, temperature, voltage, and velocity.
Extensive and intensive variables are well-definedfor many
physical and chernical systems and have been widely used. In such
systems, a component process is described by rela~onshipsbetween
its extensive and intensive variables. For example, heat flow through
a conductor is described by its heat conduction property, and the
difference in temperature is the intensive variable.Waterflow
through soil is described by soil properties and the pressure gradient in the soil. The flow of charge (or current) through a resistor
is described by the voltage drop across the resistor and its resistive
or energy-dissipative property.Waterflow
through soil-plantatmosphere systems is sometimes represented by resistances and water potential gradients. Extensive variables are measured at a point,
and intensive variables are measured across an object of interest.
By proper identification of intensive and extensive variables and
system components, a system diagram. canbe drawn that can be used
to derive the mathematical model of the system. Because of this commonality of extensive and intensive variables among systems, similar
methodologies can be used to model the results of analogous mathematical models from different systems. Koenig et al. (1967) and Martens and Allen (1969) describe in detail the use of these procedures
for physical systems. Smerage (1977) extended these concepts to a
broader class of agricultural, biological, and ecological systems.
One difficulty in modeling agricultural and biological systems
using this me tho do lo^ stems from the fact that we do not always
know the intensive variable for aparticular extensive or flow process,
or there may be several mechanisms causing flow. In this case, the
inclusion of the details of all intensive variables causing flow on a
diagram may unduly complicate the model and detract from its purposes. In other cases, our objectives for modeling the system and the
resultant scope and hierarchical detail may make it impractical to
express flow processes mecha~sticallyas related to intensive variables. For example, we h o w that the rate of flow of water through
a plant is regulated by water potential differences between the leaf
and soil and by resistances to flow along this path. However, if our

iological Processes

31

objective is to predict transpiration water use of a crop for application

to irrigation management, it is neither necessary nor desirable to include these mechanisms in the model. Including these details would
require model descriptions of water potentials of soil and plant and
associated details of plant and soil parameters that might be useful
for scientific purposes but not practical for these purposes.

So-called compartment modeling is a useful tool in conceptualizing

systems in which the primary emphasis is on the flow and storage
of system variables. Compartment models have been widely used to
schematically represent various agricultural, biological, and ecological systems (Dent and Blackie, 1979; Patten, 1973; Odum, 1973; Penning de Vries and van Laar, 1982; Keen and Spain, 1992; Leffelaar,
1993). Thisapproach is particularly useful for conceptualizing continuous systems. A mathematical model in the form of a set of firstorder differential equations describing the system structure can be
obtained directly from the compartment model diagram. Although
there are several variations in the symbols used to represent components in compartment models (Forrester, 1971; Patten, 1973), we
will use the Forrester (1971) notation because of its widespread acceptance in agricultural and biological literature. Forrester (1961)
originally developed the compartment diagrams to provide a pictorial representation of systems of equations for industrial dynamics
models.
A
is defined to be a quantity or level of a state
variable. Compartments represent the state variables in a system or
transformations of state variables. Examples of these state variables
are mass, volume, electric charge,and population. Many applications
in agricultural and biological systems relate to mass flow among various compartments in the system. A conceptual model of this type
consists of a network of compartments representing all relevant components of the system. The compartments are connected by lines that
represent the flow rates of the quantities between compartments.
Flow may also occur between the environment and compartments.
Figure 2 shows an abbreviated set of symbols from Forrester
(1961). The level represents the state variable at a point in time and
is shown by a rectangle. Sourcesand sinks represent the environment
of the system in that flow can occurfrom a source in the environment
into the system without affecting the environment. Likewise, flow

32

Compartment model symbols adapted from Forrester (1961). (a)

Level; (b) Rate; (c) Source; (d) Auxiliary Variable; (e) Pathway for Material
Flow; (Q Pathway for Information Flow.

from the system into a sink in the environment has no effect on the
environment. The rate symbol represents flow from onecompartment
to another, from a source to a compartment, or from a compartment
to a sink, Auxiliary variables are shown by circles and represent factors (inputs and parameters) that influence the rates. For example,
temperature may affect the rate of biomass accumulation in a crop
growth model and would be classified in this diagram as an auxiliary
variable. Solid lines represent pathways for material flow whereas
dashed lines represent information flow.
After the compartment model is constructed, one can write the
mathematical model representing the system using structural constraints and component descriptions. The Appendix lists variables
used in this chapter along with their ~efinitions and
units. From the
diagram itself the structural constraint or continuity can be applied
to each compartment level, or

33

where

xi= level of ith variable

dxj/dt = rate of change in the level of the ith variable
Ij,j = rate of flow into level i from source j
Oj,k= rate of flow out of level i to source k
By applying this condition to each compartment in the system, a set
of first-order differential equations is derived.

For example, consider the simple water flow system in Fig. 3a. The
volumes of water in each tank are the integrated extensive variables
of interest in the system. Equations describing this system are

Now we must describe the component flowprocesses.Sinceflow

from each tank is controlled by an orifice, we know that flow rate is
proportional to the square root of the height of the water surface
above the orifice multiplied by 2 times the gravitational constant.
Thus, we can further describe f1,2(t)and O(t) as follows.
Co~~on~nt des~~i~tions~

Since H1 and 3can be expressed in terms of Vs and V,, we can write

the complete model for this two-compartment system as (compare
with Fig. 3b):

When developing the structural and component equations as in

Eqs. (1)-(4), care should be taken to ensure the dimensional consistency of the relationships. The model developers must always make
sure that all terms have the same dimensions and that the same measurement system is used for each parameter,input, variable, and process. Otherwise, the resulting model results will not be meaningful.
Figure 3c shows an example of simulated results obtained by programming the model and simulating its behavior for 100 h, for the

FIGURE3 (a) Schematic of a two-tank water flowsystem; (b) the corresponding compartment model using Forrester (1961) notation; and (c) example results.

parameters shown in the figure. When the input to the first tank
was set to 5 m3/h for the first 5 h of simulation, then set to 0.0 for
the remaining time, the volume of water in the first tank first increased to 16.9 m3, then droppedto 0.0 after 19.1h.Volume in
the second tank lagged behind that in the first tank, peaking at only
7.1 m3.
This example demonstrates the simplicity with which this approach can beused to obtain a first-order differential equation model
of a system, Agricultural examples of compartment modeling are insect population dynamics and growth of crops. Insome such systems,
the mechanisms causing flow are not well enough understood or are
so complex that research may be required to determine relationships
between flow and other system characteristics and inputs.
In general, a model developed in this way will have state variables defined by levels xl,x,, . . . , x,. The mathematical model will
be of the form

35

Sim~~ation
of Biological Processes

20

15

10

0
0

15

10

HOURS
FIGURE3 Continued

20

25

where fl(xl,xz, . . , represents the rates as affected by any of the

variables
x2, . . . ,
and bi(t) are constant or time-varying inputs
to the ith storage compartment. For special cases when the relationships are linear, Eq. (5) can be abbreviated as
dX

-=Ax+

dt

where and b are vectors and A is a matrix of coefficients.

Cold (1982) describes another useful diagrammatic approach
called a signal-flow graph. In this approach, variables for the system,
including state variables, flow variables, inputs, and parameters, are
defined. The variables are connected by directional arrows in a diaB
C
gram to denote functional dependency. For example, A
would indicate a functional dependence of A and C on B, i.e., B =
f(A, C). This approach is very useful during the early model formulation stages to obtain an overview of the needed mathematical
relationships, such as f(A, C) in the above example, and to identify
important feedback loops that may occur in a system.

A distributed system is one in which the state variables and flow

processes vary over the spatial region of interest. In other words, the
state of the system may change through space as well as through
time. The mathe~aticalmodel that results from an analysis of a distributed system is a partial differential equation.
It is useful to break distributed systems into a set of connected,
homogeneous elements for modeling purposes. For example,we may
wish to study the flow of water and its state in a soil profile. At any
point in time, soil water content and its flow may vary with depth.
If we considered the entire profile to be a single compartment, it
would not adequately represent the vertical variations in water content with time. Therefore, consider the soil profile to be divided into
small, discrete layers, with each layer having a volume of water
stored in it. Each layer is connected to layers above and below by
flow of water; 9i,j is the volume flow rate of water from layer i to
layer j , Figure 4 shows a schematic of the system and corresponding Forrester diagram representation. From this,we can write a series
of first-order differential equations describing the structure of the
system.

37

Layer n - l

Layer n

FIGURE4 (a) Schematic of a distributed soil system (one-dimensional)for

studying soil water processes and (b) the correspondingcompartment model
using Forrester (1961) notation.

38

Jones and Luyten

Ignoring gravity effects, flow 9 can be represented by the diffusivity

property of the soil, D, and the gradient of volumetric water content,
0 (Hillel, 1971). Thus,

where fli = wi/ V = wi/Ahi and hi is the thickness of layer i.

This system of equations can be simulated if initial conditions
wi(0) and boundary conditions i(t), O(t)are known for i = 1, 2, . . . ,
n and t > 0. Initial conditions refer to the values of all state variables
at the start of the simulation, usually at time t = 0; they are required
for si~ulatingthe behavior of a model. As an example, if a system
has
state variables x(t) and y(t), then x ( 0 ) and y(0) must be specified in order to simulate system behavior for t > 0. Boundary conditions refer to values of variables at the boundaries of a system.
Boundary conditions may refer to flows across the system boundary
or to known values of intensive variables at the boundary. Values of
boundary conditions must be specified for all time t9 be simulated.
NOWwe show how this set of equations is related to a partial
differential equation. Let hi and dt be approximated by Ax and At,
respectively. Sincewi = Ah&, we can approximate dwi/dt in difference
form as

We can equate this expression with the right-hand side of Eq. (7) and
expand it into

iological Processes

Assuming L) is variable, we can write

which can be written

Taking the limit as At

0 and as Ax

0 results in

a8 a
-"
Daxax at
where 8 is a function of both x and t. Equation (13) is the partial
differential equation for water flow in one dimension through a soil
of unit area, the same equation presented by Gardner (1959). The firstorder differential equation system [Eq. (7)] is an approximate
"lumped" representation of the distributed (one-dimension)soil system more accurately represented by the partial differential equation,
Eq. (13). When approximate solutions are adequate, the lumped representation provides a relatively quick and simple method for simulating the behavior of distributed systems.
The same procedure has been used to lump age structures in
population dynamics models. Figure 5 shows the Forrester diagram
for an insect model with the population divided into age classes to
simulate age structure. The levels are population numbers for individuals in specific age classes. Anage class maybe defined as a range
of ages, for example 0-10, 10-20, 20-30, and >30 represent four age
classes of insects. Flowbetween age classes occursas insects age. The
partial differential equation describing this system is

aN(a, t ) aN(a, t )
+= -u(a, t)N(a, t )
at
aa

(14)

where
N(a, t ) = population of insects age a at time t, dimensionless
u(a, t ) = net flux of insects age a at time t across the system
boundary, S-*

u(a, t ) represents mjgration and mortality factors.

I
l
l
I
l
1

I
I
l
I

I
I
l
I
I
l
I
I

I
I
I
l

l
I
I

FIGURE5 Compartment model of insect population, with insects distributed

into four age classes.

In continuous systems, computer simulation is performed to solve

the set of system equations to result in a time series behavior of state
variables of the system for a prescribed set of inputs. This involves
one of several procedures for
advancing time in small steps and
(2) calculating the state of the system for the new point in time. This
can easily be shown for one time step using a simple model d x l d t =
ax + b. Letting dx =
- x,) and dt = ht, we can write

and
&+At

xt

+ (axt + b)At

(16)

If xi is known, we have all the information needed to calculate & + A t

or x at time t + At. It is helpful to think of this in terms of a rate,
expressed as
&+At

= xk

+ Rate, At

(17)

where Rate, is any expression for the rate of change of the state variable x at time t. In the example above, Rate, = axt + b.
Equation (16) is a finite-difference expression of the differential
equation. In general, Eq. (16) is also referred to as the rectangular or
Euler method of integration of a first-order differential equation. The
value of x is known at time t and possibly prior to time t. The model
expresses the instantaneous rate of change of the variable x. Although
we do not know the value of x at t + At, we can choose a small
and predict a value of x at time t + At based on the current rate of
change of x. This procedure is repeated as long as desired. For example, by continuing with x at t + At and projecting forward by one
more At we can estimate x for time t + M t .
The result will be a time series behavior of x defined by the
model. If the system has more than one first-order differential equation, the finite difference procedure can be applied to each. The rate
expression for any one differential equation may be functionally dependent on any of the state variables as well as on inputs to the
system. This method is dependent on the system model being described by first-order differential equations (or difference equations
as described earlier). Second-order differential equations can be simulated by transformation to two first-order equations. For example,
the second-order equation

can be transformed into two first-order equations by letting x1 = y

and xz = dyldt. It follows that dx,/dt = d2y/dt2.Then the two equivalent first-order differential equations are
dx,
= x2
dt

and

dx2
dt

-=

ax2 - x1 + u(t)

Other numerical solution techniques are more accurate than the

Euler method for the same size At. For example, the trapezoid integration method can be used to develop a "predictor-corrector" integration technique. In this case,

is the formula for calculating yf+At. Note that Rate:+t is on the righthand side of Eq. (20). We have to "predict" Rate:+,, since we do not
h o w y or Rate at time t + At. The Euler method is used to predict
is used to compute
as a first approximation. Then that Y:+,~
Rate;',,, using the first-order differential equation model. We then
have all that is needed to compute Y f + A t above. The Euler and trapezoid methods are ex~Zicitmethods; all terms on the right-hand side
of the equation are h o w n or can be computed from what is known.
In general, a Taylor's series expansion about yt can be used to
express y;",,. This expression is

where
represents the nth derivative of y with respect to t and n!
represents factorial numbers (i.e., n! = 1 2
4 5
- n).
If we approximate yt by using the first three terms of Eq. (2l), then

where c t is the total truncation error, or the error caused by not including all the terms in the Taylor expansion.
Since
can be approximated by
- y',")/ At, Eq. (22) can
be written
Yt+At

= Yt

Et

Equation (23) is equivalent to the trapezoid predictor-corrector technique with the added error term. The Euler technique is equivalent
to the Taylor series expansion with only the first derivative term included explicitly.Thus, the trapezoid method is a more accurate
method for simulating first-order differential equations. However, the
Euler method is used for most models because of its simplicity and
ease of use.

iological Processes

43

Because that is a numerical approximation, it is subject to numerical errors. The local truncation error using the Euler method is
[(At)z/2][dzy(~)/dtz],
where 6 is some point in the interval [t, t + At].
The truncation error is the error with each integration step and
can thus accumulate. Truncation error for the trapezoid predictorwhich, for a small At, is less
corrector method is [(At)3/6][d3y(~)/dt3],
than that of Euler method (Ortega and Poole, 1981). Generally the
higher order methods are more stable than lower order methods.

The purpose of sensitivity analysis is to study the behavior of the

model. Sensitivity analysis can also be structured to determine important subsystems, relationships, and inputs. Sensitivity analysis
should be designed with respect to the objectives of the study. A
sensitivity analysis is the process by which parameters or inputs are
evaluated with regard to their effects on simulated results. For example, rainfall would be an environmental input to a crop balance
model for estimating crop irrigation requirements. Soil characteristics
such as soil water-holding capacity or root zone depth are examples
of model parameters. One may be interested in the sensitivity of estimated irrigation requirements to changes in rainfall, changes in soil
characteristics, or both. A sensitivity analysis also provides a mechanism for testing the simulation in the extremes; that is, using the
extreme values of parameters will rigorously test the model in terms
of mathematical logic and stability.
Computer graphics are highly useful in sensitivity analysis.
Graphical display of simulated results for a number of parameter
values or inputs provides a visual image of model behavior over a
range of parameter values. A mathematical approach to sensitivity
analysis may also be a useful way to organize and present model
behavior. This approach compares the change in one or more simulated outputs relative to the change in one or more parameters by
approximat~gpartial derivatives using numerical results. Absolute
sensitivity a(y I k), of some model output y to a variable k (a parameter
of the model or input to the system) is defined by

Relative sensitivity a,(y I is often used to provide a normalked measure for comparing the sensitivity of a model to several variables. Relative sensitivity is defined by

For example, if cr,(y I k ) = 2.0 and k is changed by 30% (

we can expect to change by
Also, if a,(yl k,) = 0.5 and o,(y I kz)
= 1.5, then y is more sensitive to a percentage change in kz than the
same percentage change in k, (3 times as sensitive).
In simulation studies h k is used to represent a small change in
k so that changes in y,Ay, can be simulated. Then Eqs. (24 and (25)
can be used to approximate cr(y I k ) and cr,(y I k ) using the discrete
changes in k and y, or dy/ak
~ y / h k .By varying parameters
through their expected ranges of extremes, one can use this approach
to compare the sensitivity of model results.
Sensitivity analysis usually begins with the selection of model
output results that are considered to be crucial to the study. A set of
base conditions are then established; these usually comprise the set
of the best estimates of each parameter and input. Base results are
the simulation,outputsobtained when base condition values are used.
A range of values are then selected representing the extreme conditions associated with each parameter to be evaluated. Simulation runs
are made using each value while holding all other base conditions
constant. A comparison between changes in the base condition values
and changes in the base results provides an indication of the relative
importance of the variable. Results can be compared in graphs or
tables or by computing the sensitivity variable values. Note that the
values of cr and
and graphical results depend on the base conditions selected.
Regression analysis or response surface techniques are also useful techniques in sensitivity analysis. Computer experiments are conducted by varying parameters over a range of interest followed by
regressing simulated results against one or more parameters to determine if outputs are affected by changes in the parameter, and if
so, to what extent.

.
There are many options for computer implementation of biological
models today. Becausethere are many reasons for developing biological models, some consideration should be given to which option to
choose for a particular model. Some biologicalmodels are developed
by an individual or small team of scientists to test a hypothesis, and
there is no attempt to make these models available to others except

through the scientific literature. Other models are developed by teams

of researchers foruse in a number of applications and for distribution
to others who may want to use them for their own purposes. Biological models are also being used in formal university classrooms and
in specialized short courses and training programs. Programs should
be constructed using modules or objects designed for describing welldefined components of the system being modeled and for ease of
documentation, revisions, replacement, and maintenance. Van Kraalingen (1995) describes a structure for programming in Fortran that
provides these characteristics. In many applications, biological models need to be linked with other software, such as geographical information systems (e.g., Engel et al.,1996;Fraisse et al., 1994) and
graphical user interface programs (e.g., Jacobson and Jones, 1996).
Standardized formats for data storage and exchange are needed to
facilitate model evaluation and tolink models with such applications.
Some progress has been made on standardizing data for cropmodels
and their applications (Jones et al.,1994); this effort is continuing
at an international scale (Ritchie, 1995). However,considerably more
effort is needed by biologicalscience communities on standard
protocols for biological model module design and data storage and
exchange.
There are many choices of programming languages and software
to help those who are developing biological models. The choice of
language is not a major issue today because of the ability to link
together modules developed in different languages,particularly if the
models use modular structure and adhere to well-defined data standards. However, each has its own advantages and disadvantages.
Procedural languages such as Fortran, Basic, and Pascal are widely
used for biological models. Most existing crop models, for example,
are programmed in Fortran (Jones and Ritchie, 1990; Hoogenboom
et al., 1994; Bouman et al., 1996). Procedural languages provide excellent capabilities for scientificcomputations, they can easily handle
nonlinearities and other complexities commonto many biological systems, and they can be used on many computer platforms with little
or no modification. Because of limitations in these languages with
respect to handling modern user interface programming, new applications of these models are being developed in which the
model is treated as a module in an overall software package and the
user interface and data manipulations are handled by one of the
modern visual programming tools such as Visual Basic micros so^,
1995). Some biologicalmodels are now being developed using objectoriented programming languages such as SMALLTALK and G+ +.

Jones and luyten

These new languages have potential for helping improve the modularity of biological models.
There are a number of specialized computer simulation languages that allow model developers with little or on computer programming skills to implement models. Continuous simulation languages solve systems of continuous differential equations. For
example, the SLAM system (Pritsker, 1995) and FSE (van Kraalingen,
1995) packages provide modules that perform most of the required
computer simulation tasks. Users only have to "program" their specific model equations and provide parameters and other inputs in
order to obtain simulated behavior of a system. Probably the simplest
and easiest to use simulation software allows users to build conceptual models on computer screens using icons, such as Forrester symbols presented earlier in this chapter. Two widely used packages are
Stella (High Performance Systems, 1990) and VisSim(VisualSolutions,1990), both available for use on personal. computers. These
packages guide users in linking together different icons to depict a
system with its storage compartments and flow paths, then ask for
needed parameters and other inputs to run the
model, They willthen
simulate the system and graph or print results for users; no programming is needed. These packages have the distinct advantage of rapid
evaluation of variations in model structure to test different hypotheses about the system. Each computer language has its advantages
and disadvantages. The best choice for a particular model will depend on many factors such as model complexity, intended users of
the model, how it will be used, the need to integrate it with other
software and databases, and the need to maintain it.

Dynamic growth and yield models have been developed for a wide
range of crops, including cotton, grain sorghum, wheat, corn, soybeans, alfalfa, and others (Jones and Ritchie, 1990). Researchershave
used these models for a number of purposes, such as irrigation management (Boggess et al., 1981; Swaney et al., 1983), nutrient management (Keating et al.,1993; Singh et al. 1993), pest management
(Pinnschmidt et al., 1990; Batchelor et al., 1993), land use planning
(Beinroth et al., 1997),crop sequencing (Thornton et al., 1995),climate
change assessment (Curry et al., 1995; Rosenzweig et al., 1995; Penning de Vries, 1993), and yield forecasting (Swaney et al., 1983). These
applications have demonstrated conclusively the value of simulation
applied to cropping systems. However, these models have limitations

S i m ~ ~ a t of
i o Biological
~
Processes

47

because they do not include all factors that occur in reality and
contain empiricism that may require calibration and testing for sitespecific applications. Research directed to improving these shortcomings and limitations is needed to enhance the applicability of the crop
models.
A useful characterization of crop models for application at the
field scalewas attributed to C. T. de Wit (Bouman et al., 1996;Penning
de Vries and van Laar, 1982). Initially, four levels were defined, but
these were later simplified to three levels (Lovenstein et al., 1993).
First is the potential yield level; crop growth is dependent onweather
factors, such as temperature, solar radiation, CO, concentration, and
day length, and on genetic factors of the crop. This level of model
includes basic crop growth processes such as photosynthesis, respiration, tissue growth, and development. Its main use is to gain an
understanding of how these factors affect potential production of a
crop, assuming that water and fertilizers are adequate and that no
pest damage occurs. A second level of detail is characterized as attainable yield; at this level, the model has components that describe
water and nitrogen fertilizer limitations to production. Attainable
yield thus depends on water and nitrogen supplied by management
as well as by rainfall and soil organic matter. Many existing crop
models have the capability to describe crop growth and yield with
these assumptions (Jones, 1993).The third level of detail is actual
yield level, which attempts to explain all the factors that influence
growth in a field, including pest damage and mismanagement. This
characterization of levels of detail has been useful in providing both
a practical pathway for model development and a conceptual framework for model applications. It has also served to determine data
requirements for model development, taking into consideration the
need for more comprehensive data sets as the level of model detail
increases.
There are many choices of state variables for crop models. Simple models may have only one or two state variables, such as leaf
area and dry weight, whereas others may have hundreds of state
variables describing leaf area, leaf weight, number of leaves, stem
weight, number of fruit, and weight of fruit for various age classes
(Jones et al., 1991). Crop models with many state variables provide
more details about crop behavior, including lags in fruit growth and
size distributions, but require much more time to develop and test
and more time to solve in a computer. For some problems, models
with two to four state variables provide adequate predictions of crop
responses. In this chapter, a simple crop model at the potential yield

~ l l o t ~ ~ t h ~ l s
param~e~

h h t (I.)

temperature (7"

.""""""":""""""""""-"-

""""

IGURE 6 Forrester diagram for the crop model with three state variables
at the potential productionlevel of detail.

level of detail is presented along with its sensitivity to temperature

and light.
Figure 6 shows a Forrester diagram of a simple crop model with
three state variables: stage of development (M,number of vegetative
nodes), canopy biomass (W,, g/m2), and root biomass (Wr, g/m2).The
diagram represents a potential yield level crop model as described
by Lovenstein et al.
The figure also shows a compartment for
assimilates (Ap, g CH20/m). However, in the mathematical model, it
is assumed that d A , / d t = 0.0, and thus the assimilates do not accumulate but are transformed directly into plant tissue.

ent" refers to the progression of the crop through its life

stages, which may be measured by various vegetative and reproduc-

tive plant characteristics such as number of leaves on the plant or the

appearance of the first flower or first fruit. Development rate is usuas the inverse of time between events, such as the
rs between successive leaves on the main stem or of
days from plant emergence to flowering. Theserates of development
can thus be calculated from experiments in which stages or numbers
of leaves are recorded. Development rates in plants are very sensitive
to temperature and sometimes to day length but are usu
sitive to light, COz, and other factors such as water a
stresses unless the stresses are severe. Crop developme
insensitive to rate of dry matter production.
There have been many models for predicting crop development; most have related development rate to temperatureand some,
to day length. For example, Wolf et al. (1986) developed a model for
scheduling field operations that predicted emergence,flow
and harvest dates for field tomatoes within
days of actual
et al. (1991) described a system for predicting development
for cucumbers, eggplant, and other crops growing in greenhouses. These models are mostly of the form Y = r(T, q), where Y is
the rate of development (h-'), T is temperature, and q is day length.
In models for predicting flowering data, Y is the rate of progression
toward flowering, and the state variable in these models is the development unit, which is 0.0 when the development starts an
at flowering.
For the three-state variable model, N is the state variable for leaf
number for a vegetative ant or for node number for plants 1
tomatoes that alternately
duce leaves and f ~ i t i n g
trusses on the
main axis. Assuming that leaf appearance rate, d N l d t , depends only
on temperature in this example, it can be computed as
dN
- = Y,Y(T)
dt

where Y, is the maximum rate of leaf appearance (h-') and r ( T ) is a

function of temperature. Various functional forms have been used for
~ ( 7 ' )In
. the model presented here, a piecewise-linear form was used
on the tomato model (Jones et al., 1991), where r ( T )
is 0 when T is below 8C or above 5O"C, and has the value
1.0, and 1.0 at temperatures of
and 35"G, respectively.
30 and 35C for this example, leaf appearance rate is maximum at Y,
leaves per hour.

Jones and Luyten

The rate of dry matter increase in crops is of major importance in

determining yield of fruit and vegetative tissue, and it depends directly on photosynthesis. Photosynthesis converts carbon dioxide
from the air into sucrose (CH,O) in plant leaves. Some of the carbon
in this CH20 is combined with other elements (N, ]p, K, S, . . .) and
retained in plant tissue, and some is used to provide energy for synthesizing tissue, releasing CO, in the process. Penning de Vries et al.
(1989) describe a sound quantitative basis for computing the rates of
tissue synthesis based on the photosynthetic rate and on the composition of tissue being synthesized. They computed the net product
weights of protein, carbohydrate, lipid, lignin, organic acid, or mineral if l g of CH,O was used to provide both the carbon in the final
product and the energy for product synthesis. Thus, depending on
E, can be comthe composition of plant tissue, a conversion efficiency;
puted and used to convert photosynthesis rate into dry matter accumulation rate. For vegetative tissue, this conversion efficiency is in
the range of 0.65-0.75 (g tissue)/(g CH,O).
Respiration is the loss of GO, from plants in growth and maintenance processes. Growth respiration is accounted for in the conversion efficiency described above. ~aintenancerespiration is the loss
of CO, due to the breakdown and resynthesis of existing tissue and
depends on temperature. Gent and Enoch (1983) expressed maintenance respiration rate as
R, = k , expj0.0693(T
(27)- 25)]
where

R, = maintenance respiration rate, (g GH20)/ [(g tissue) h]

T = temperature, "G
= respiration rate at 25"C, (g GH20)/[(g tissue) h]
The equation for crop dry weight growth rate is
dW

- = E(P,
dt

R,W)

where

~ W / d=trate of dry weight growth of crop, (g tissue)/ (m2 h)

W = total plant dry weight, g/m2
R , = maintenance respiration rate, (g CH,O) / [(g tissue) h]
E = conversion efficiency ofGH,O to plant tissue, (g tissue)/ (g CH20)
P, = canopy gross photosynthesis rate, (g CHzO)/(m2 h)

51

Total biomass growth rate is partitioned into canopy and root

biomass with the function f,(N), or dW, /dt = ( d W / d t ) f c ( ~ and
) , dWr/
dt = (dW/dt)[l.O - fc(N)]. This partitioning of new growth between
canopy and root varies with stage of development; however, in this
example, we will assume that fc
An
expression is needed to
opy gross photosynthesis
rate, P,, as affected by light, CO,, temperature, and plant size. There
are many models that describe leaf and canopy photosynthesis;
Charles-Edwards (1981) gives the mathematics of photosynthesis processes in detail. The model of Acock et al. (1978) was found to adequately describe tomato canopy photosynthesis rates (Jones et al.,
1991). This equation, modified to include temperature effects, is

where

D = coefficient to convert photosynthesis calculations from

(pmol CO,) / (m2* S) to (g GH20)/ (m2 h)
I = leaf conductance to CO,, (pmol CO,) / [(m leaf) S]
C = CO, concentration of the air, (pmol CO,) / (mol air)
p ( T ) = photosynthesis reduction factor, dimensionless
a = leaf light utilization efficiency, (pm01 CO,)/ (pm01 photons)
K = canopy light extinction coefficient, dimensionless
Io= light flux density at the top of the canopy (pmol photons)/ [(m ground) * S]
m = light transmission coefficient of leaves, dimensionless
L = canopy leaf area index, (m leaf)/ (m2ground)
The function p() expresses the effect of temperature on the maximum rate of photosynthesis for a single leaf, expressed as a quadratic
equation with T :

where +h is the temperature at which leaf photosynthesis is maxim~m

and +l is the temperature below which leaf photosynthesis is zero.
Hourly values of I, were computed using the method of Goudriaan
(1986), assuming
h day lengths (i.e.,
th 18):

I, = Imsin{Zn[( t , -

/ 241)

(31)

52

where I,nis the maximum light flux density (at noon) and th is the
solar time in hours,
Equation (29)contains environmental variables (T, C, Io),various
parameters a,K, m, D), and thecanopy leaf area, L, which depends
on time. Leaf area ratio (square meters of leaf area per gram of plant)
and specific leaf area, (m2leaf)/ (g leaf), vary considerably with env~onmentalconditions. Under cool temperatures, leaves appear more
slowly [Eq. (26)], accumulate more dry weight during the extended
growth period, but expand to about the same final area over a practical temperature range, as shown by Jones et al. (1991) for tomato.
Under low light, leaves will tend to be thinner because of lower rates
of photosynt~esisper unit of development. The assumption that L is
a function of N provides a model that mimics observed leaf area ratio
responses to light, temperature, and CO,. The expolinear equation
(~oudriaanand ~ o n t e i t h ,1990) was used to fit leaf area vs. node
number using data reported by Jones et al. (1991) for tomatoes grown
at two CO, levels and three temperatures, The equation is
L =
where

P ) W + exp[P(N - %)l>

(32)

L = leaf area index, (m2leaf) / (m2ground)

p = plant density number / m2
N = leaf number
and S, and nb are empirical coefficients forthe expolinear equation.
Therefore, the model predicts the rate of leaf appearance, which
primarily depends ontemperature, and calculates leaf area. Thefinal
three-state variable model can be expressed as

dN

-= rn,r(T)
dt

(33)

qs. (27) and (29)-(32) required to compute R,, PS,

L, respectively. Theseequations could be substituted into
this is not necessary because computer simulation is U
the equations. For this study, a time step ( t) of l h was used, and
the model was simulated for 80 days. For each simulation, environmental inputs (C, T, and Io)were held constant for each run, and a
number of combinations were simulated to de~onstratemodel behavior. Table l shows the values for coe~cientsused in this model,
most of which were taken rom Jones et al. (1991).

le 1. Values of Coefficients Used in the Crop Model Example

= 0.056

p = 0.38
= 0.074
+h
$1

=I

30.0

= 5.0

= 0.0664
= 4.0
k , = 0.0006
m = 0.10
n b = 13.3
T

r, = 0.021
C = 350
D = 0.108
E = 0.70
fc(N) = 0.85

1, = 11200
K = 0.58

There are limitations in the use of this model. Itdoes not account
for senescence or picking of plant material, nor does it partition dry
matter into different plant components, such as fruit. The model
could be extended to describe the growth of fruit and other organs,
thereby creating more state variables, equations, and a need for more
coefficients and relationships.

Figures 7a and
show the effects of temperature, light, and CO,
concentration on the rates of photosynthesis predicted by Eq. (29)and
the effects of temperature on respiration for a canopy with a leaf area
index ( L ) of 4.0. These results demonstrate the importance of weather
on the rate of dry matter accumula~on.Figure 7c shows the effect of
temperature on leaf area expansion over time. Low temperature reduces the rate of node formation [Eq. (26)], which results in delays
in leaf area development. When plants are young, rapid leaf area
expansion is important to create a h11 canopy for capturin~light for
photosynthesis.
The three-state variable model, Eq. (33), was simulated for 80
days under assumed constant environmental conditions to demonstrate the integrated and cumulative effects of growth and de
ment rates on yield. For each set of ~omputations,all variab
one were held at reference values and one variable was chan
successive runs. Reference conditions were 30/20"C day/ night temperatures, 12 h days, 1200 (pmol photons)/ (m2 S) maximum light
flux density, and 350 (pmol CO,) / (mol air).
Figure 7d de~onstratesa major effect of temperature on
weight over the 80 day simulations. Temperature affects both development and photos~thesis,but its major effect is on development.
For the l8 / 8C temperature case, 59days were required to reach node
l 0 in contrast at 23 days for the 30/20"C te~peraturecase. ~hotosynthesis for the 18/8"C case was reduced by only about 30%. For this

500
l000
1500
2000
LIGHT FLUX DENSITY(~mol(photon)/mA2-s)

F I ~ U R E7 Simulated results from the three state variable crop model. (a)
Photosynthesis VS. maximum light flux (I,) at four CO, levels at temperature
ln =: 30C; (b) photosynthesis vs temperature at four maximum light flux
levels for a canopy with leaf area index L = 4.0; (c) leaf area development
vs time for four temperatures; (d) total dry weight vs. time for four day/
night temperatures; and (e) total dryweight vs. time for fourlight levels at
3O/2O0C day /night temperatures.

coldest case, dry matter at $0 days was lower by a factor of 15. Note
that growth was lower at 38/28"C thanat 30/20"C day/night
temperatures.
As light was varied from 400 to 1600 (pm01 photons)/ (m' S),
dry weight yield increased by a factor of2.5 at 80 days (Fig. 7e).
Numbers leaves and leaf area were not affected by light, Increasing
GO, levels from an ambient level 350 (pmol CO,) / (mol air) to 700
resulted in a 28% increase in dry weight yield at 80 days (data not
shown).
LES

Ai
Ci
D

area of water tank m,

constant,
m2 (water t a d case)
d
~
sof the
i soil,
~ m/~S

55

-l

LIGHT FLUX LEVELS

l600

IGURE 7

10

10

15
20
25
TE~PE~TURE
(c)
,

20

Continued

30

40
DAYS

50

30

60

35

70

40

80

10

20

30

40
DAYS

50

60

l
LIGHT FLUX

10
IGURE

20

7 Continued

30

40
DAYS

50

60

70

80

57

flow rate from tank i to tank j , m3/S

gravity constant, m/ S'
thickness of soil layer i, m
height of water in tank i, m
rate of flow into level i from source
population of insects age a at time t.,number
rate of flow out of level i to source
flow rate of water from layer i to layer j , m3/S
net flux of insects age a at time t across system boundary
S-l

volume in tank i, m3
volume of water stored in layer i, m3
level of ith state variable, various units
volumetric water content, m3/ m3
absolute sensitivity of output y to input k, dimensionless
relative sensitivity of output y to input k, dimensionless

assimilates, (g CH20)/ m
CO, concentration of the air, (pm01 CO,)! (mol air)
photosynthesis conversion
coefficient,
(g C
CO,)
conversion efficiency at CH20 to plant tissue, (g tissue)/(g
CHZO)
fraction of total crop growth partitioned to canopy dimensionless
actual light flux density at top of canopy (pm01 photons)/
(mZ S)
maximum light flux density at top of canopy (pm01 photons)/ (m2 S)
respiration rate at 25"C, (g CH20)/ [(g tissue) h]
canopy light extinction coefficient, dimensionless
canopy leaf area index, (mZleaf) / (m2ground)
light transmission coefficient of leaves, dimensionless
empirical coefficient for expolinear equation, ~imensionless
leaf number or node number, dimensionless
photosynthesis reduction factor, dependent on T, dimensionless
canopy gross photosynthesis rate, (g C
rate of development, h-'
maximum rate of leaf appearance, h-'
+

maintenance respiration rate, (g CH20)/ [(g tissue) h]

tem~erature,"C
solar time, h
total plant dry matter weight, g/m2
canopy dry matter weight, g/m2
root dry matter weight, g/m2
light utilization efficiency (pmol CO,) / (pmol photons)
empirical coefficient for expolinear equation, dimensionless
empirical coefficient for expolinear equation, dimensionless
day length, h
plant density; m-2
leaf conductance to CO2, (pmol CO,) / [(m2leaf) - S]
temperature at which leaf photosynthesis is maximal, "C
temperature at which leaf photosynthesis is zero, "C

Acock, B., D. A. Charles-Edwards, D. J. Fitter, D. W. Hand, L. J. Ludwig, J.

Warren Wilson, and A. C. Withers. 1978. The contribution of leaves from
different levels within a tomato crop to canopy net photosynthesis: An
experimental examination of two canopy models. J, Exp. Bot. 29:825-827.
Batchelor, W. D., J. W. Jones, K. J. Boote, and H. Pinnschmidt. 1993.Extending
the use of crop models to study pest damage. Trans. A S A E 36(2):551-558.
Beinroth, F. H., Jones, J. W., E. B. Knapp, P. Papajorgji, and J. C. Luyten. 1997.
Application of DSSAT to the evaluation of land resources. In: ~ n ~ e r s t a n ~ ing ~ t i o n for
s Agricultura2 ~ r o ~ u c t i oG.
n . Y. Tsuji, G. Hoogenboom, and P.
K. Thornton (Eds.). Dordrecht: Kluwer Academic. (in press).
Boggess, W. G., J. W. Jones, D. P. Swaney; and G. D. Lynne. 1981. Evaluating
irrigation strategies in soybeans: A simulation approach. In: Proceedings
of the ASAE irrigation Scheduling Conference, Dec. 14-15, 1981. St.
Joseph, MI: ASAE, pp. 45-53.
Boote, K. J., J. W. Jones, and N. B. Pickering. 1996. Potential uses and limitations of crop models. Agron. J. 88(5):704-716.
Bouman, B. A. M., H. van Keulen, H. H. van Laar, and R. Rabbinge. 1996.
The "School of de Wit" crop growth simulation models: A pedigree and
historical overview. Agric. Syst. 52(2-3):171-198.
cs
and ~ r o d u c t i ~ Charles-Edwards, D.A. 1981. The ~ a t h e ~ ~ tofi ~hotosyn~hesis
ify. New York Academic Press.
Curv, G. L., and R. M. Feldman. 1987. ~ a t ~ e m a t i c a l ~ o u n ~oft i~oon~su l a t i o n
~ y n a ~ i cTEES
s . Monograph Series, No. 4. College Station,
TX Texas A&M
Univ. Press.
Curry, R. B., J. W. Jones, K. J. Boote, R. M. Peart, L. H, Allen, Jr., and N. B.
Pickering. 1995. Response of soybean to predicted climate change in the
USA. In: Climate Change and Agriculture: Analysis of Potentia2 ~~ternational

Impacts. C. Rosenzweig, J. W. Jones, G. Y. Tsuji, and P. Hildebrand (Eds.).

ASA Special Publication No. 59. Madison,
WI: American Societyof Agronomy pp. 163-182.
Dent, J. B., and M. J. Blackie. 1979.Systems Simulation in A ~ i c u l t u r eLondon:
.
Applied Science.
Engel, T., J. W. Jones, and G. Hoogenboom. 1996. AEGIS/ WIN-A powerful
tool for the visualization and comparison of crop simulation results. In:
Proceedings of the Sixth Internation~ZConference on Computers in A ~ i c u l t u r e .
F. S. Zazueta (Ed.). Cancun,Mexico, June l-14,1996. ASAE Pub. 701P0396.
St. Joseph, MI: ASAE, pp. 605-612.
Forrester, J. W. 1961. Industrial Dynamics. New York Wiley.
Forrester, J. W. 1971. Princi~lesof Systems. Cambridge, MA: Wright-Allen.
Fraisse, C. W., K. L.Campbell, J. W. Jones, W. G. Boggess, and B. Negahban.
1994. Integration of GIS and hydrologic models for nutrient management
planning. Proceedings of the National Conference on Environmental Problem Solving with Geographical Information Systems. Cincinnati, OH. Sept.
21-23, 1994. EP~/625/R-95/004,pp. 283-291.
France, J., and J. H. M. Thornley. 1984. MathematicalModels in A~iculture.
London: Butteworths.
Gardner, W. R. 1959. Solutions of the flow equation for the drying of soil and
other porous media. Soil Sci. Soc. Proc, 23:183-187.
Gent, M.P. N., and H. Z. Enoch. 1983. Temperaturedependence of vegetative
growth and dark respiration: a mathematical model. PlantPhysiol. 71:
562-567.
Gold, H, J. 1982. Mathe~aticalModeling of ~ i o Z o ~Systems:
c~l
An Introducto~
Guidebook. New York Wiley.
Goudriaan, J. 1986. A simple and fast numerical method for the computation
of daily totals of crop photosynthesis. Aguic. Forest Meteorol. 38:251-255.
Goudriaan, J. 1988. The bare bones of leaf-angle distribution in radiation
models for canopy photosynthesis and energy exchange. Aguic. Forest Meteorol. 43:155-169.
Goudriaan, J., and H. H. van Laar. 1994. Modelling Potential Crop Growth Processes: Textbook with Exercises. Current Issues in Production Ecology, Vol. 2.
Dordrecht: Kluwer.
Goudriaan, J., and J. L. Monteith. 1990. Mathematical function for crop
growth based on light interception and leaf area expansion. Ann. Bot. 66:
695-701.
High Performance Systems. 1990. Stella II Users Guide. Hanover, NH: High
Performance Systems.
Hillel, D. 1971. Soiland Water PhysicalPrinciplesandProcesses.
New York
Academic Press.
Hoogenboom, G., J. W. Jones, P. W. Wilkens, W. D. Batchelor, W.T. Bowen,
L. A. Hunt, N. B. Pickering, U. Singh, D. C. Godwin, B. Bear, K. J. Boote,
J. T. Ntchie, and J. W, White. 1994. Crop models. In: DSSAT v3.Vol.
2.G. Y. Tsuji,G. Uehara, and S. Balas(Eds.). Honolulu: University of
Hawaii, pp. 95-244.

Jacobson, B. M,, and J. W. Jones, 1996. Designing decision support system

for soybean management. In: Proceedings of the Sixth ~nternatio~al
Conference
on Computers in A ~ i c u l t u r e .F. S. Zazueta (Ed.). Cancun, Mexico, June
14, 1996. ASAE Pub. 701P0396, St. Joseph, MI: ASAE, pp. 394-403.
Jones, J. W. 1993. Decisionsupport systems for agricultural development. In:
Systems Approaches for A ~ i c u l f u r a lDevelopment. F. Penning de Vries, P.
Teng, and K. Metselaar (Eds.).Dordrecht: Kluwer Academic, pp. 459-471.
Jones, J. W., and J. T. Ritchie. 1990. Crop growth models. In: Management of
Farm ~rrigation Systems.G. J. Hoffman, T.A. Howell, and K. H. Solomon
(Eds.). St. Joseph, MI: Am Soc. Agric. Eng., pp. 61-89
Jones, J. W., E. Dayan, L. H. Allen, W. van Keulen, and H. Challa. 1991. A
dynamic tomato growth and yield model (TOMGRO). Trans. ASAE
663-672.
Jones, J. W., L. A. Hunt, C. Hoogenboorn, D. C. Godwin, U. Singh,G. Y. Tsuji,
N. B. Pickering, P. K. Thornton, W. T. Bowen, K. J. Boote, and J. Ritchie.
1994. Input and output files. In: DSSAT v3. Vol. 2. G. Y. Tsuji, G. Uehara,
and S. Balas (Eds.). Honolulu: University of Hawaii, pp. 1-94.
Keating, B. A., R. L. McCown, and B. M. Wafula. 1993. Adjustment of nitrogen inputs in response to a seasonal forecast in a region of high climatic
risk. In: Systems Approaches for A~icultural Develo~ment.E Penning de
Vries, P. Teng, and K. Metselaar (Eds.). Dordrecht: Kluwer Academic, pp.
233-252.
Keen, R. E., and J. D. Spain. 1992. C o m ~ u t e r S ~ ~ u l a t i o nBiology:
in
A Basic
~~troduction,
New York, Wiley-Liss.
r, C. A., J. W. Jones, and E. R. Muller. 1991. A model and database for
predicting phenomenological stages. ASAE Paper No. 91-7035. St. Joseph,
MI:ASAE.
Koenig, H. E., Y. Tokad, and D. R. Allen. 1967. Analysis of Discrete Physical
Systems. New York McGraw-Hill.
Leffelaar, P. A. (Ed.). 1993. On Systems Analysis and Simul~tionof Ecological
Processes, with Examples in CSMP. Current Issues in Production Ecology.
Vol. 1. Dordrecht: Huwer Academic.
Lovenstein, H., E. A, Lantinga, R. Rabbinge, and H. van Keulen. 1993. Principles of ~eoretical ~roduction
Ecology. Course book. Department of Theoretical Production Ecology.Wageningen,The Netherlands: Wageningen
Agricultural University
Martens, H. R., and D. R. Allen. 1969.~ntroduction to SystemsTheory. Columbus, OH: Merrill.
Microsoft. 1995. Microsoft Visual Basic, Pro~ammersGuide. Redmond, WA:
Mcrosoft Corporation.
Monteith, J. L. 1990. Conservative behaviour in response of crops to water
and light. In: he ore tical Production Ecology: ~ e ~ e c t i o and
n s Prospects. R. Rabbinge, J. Goudriaan, H. van Keulen, F. W. T. Penning de Vries and H. H.
van Laar(Eds.). Simulation Monographs 34.Wageningen,The Netherlands: PUDOC, pp. 3-16.

Qdum, H. T. 1973. An energy circuit language for ecological and social systems. In Systems Analysis and Simulation in Ecology. Vol.11. B. C. Patten
(Ed.). New York Academic Press, pp. 140-211.
Qrtega, J. M., and W. G. Poole, Jr. 1981.An Introduction to ~umericalet hods
for Di~erential E~uations.
Marshfield, MA: Pitman.
Patten, B. C. 1973. Systems Analysis and Si~ulation inEcology. Vol.11. New
York Academic Press.
Penning de Vries, F. W.T. 1993.Rice production and climate change. In:
Systems Approaches for Agricultural Developme~t.F. Penning de Vries, P.
Teng, and K. Metselaar (Eds.). Dordrecht, The Netherlands: Kluwer Academic, pp. 175-189.
Penning de Vries, F. W.T., and H. H. van Laar (Eds.). 1982. Simulation of
Plant Growth and Crop Production. Wageningen, The Netherlands: PUDQC.
Penning de Vkies, F. W. T., D. M. Jansen, H. F. M. ten Berge, and A. Bakema.
1989. Simulation ofEcophysiologicalProcesses
of Growth in Several Annual
Crops. Simulation Monographs 29. Wageningen, TheNetherlands: PUDQC.
Pinnschmidt, H., P. S. Teng, and J. E. Yuen. 1990. Pest effectson crop growth
and yield. InProceedings of the ~ o r ~ s h on
o pMode~ingPest-Crop Interactions.
P. Teng and H. Yuen (Eds.). Honolulu: University of Hawaii.
Pritsker, A. A. B. 1995. Introduction to Simulation and S L A M II. 4th ed, New
York Wiley.
Rabbinge, R., P. A. Leffelaar, and H. C. van Latesteijn.1994.Therole
of
systems analysis as an instrument in policy making and resource management. In Opportunities, Use, and Transfer of Systems Research Met~ods in Agriculture to Developing Countries. P. R. Goldsworthy and F. W. T. Penning
de Vries (Eds.). Dordrecht, The Netherlands: Kluwer Academic, pp. 6779.
Ritchie, J. T. 1995. International Consortium for Agricultural Systems Applications (ICASA): Establishment and purpose, Agric. Syst. 49:329-335.
Rosenzweig, C., J. W. Jones, G. Y. Tsuji, and P. Hildebrand (Eds.). Climate
Change and A ~ i c u l t u r eAnalysis
:
of Potential International Impacts. ASA Spec.
Publ. No. 59. Madison, WI: American Society of Agronomy.
Sinclair, T.R. 1990. Nitrogen influence on the physiology of crop yield. In:
heo ore tical Production Ecology: Reflections andProspects.
R. Rabbinge, J.
Goudriaan, H. van Keulen, F. W.T. Penning de Vries, and H. 3. van Laar
(Eds.).Simulation Monographs 34.Wageningen,The
Netherlands: PUDOC, pp. 41-55.
Singh, U,, P. K. Thornton, A. R. Saka, and J. B. Dent. 1993. Maize modeling
in Malawi: A tool for soil fertility research and development. In: Systems
Approaches for Agricultural Development. F. Penning de Vries, P. Teng, and
K. Metselaar (Eds.). Dordrecht, The Netherlands: Kluwer Academic, pp,
253-273.
Smerage, G. H. 1977. On the extension ofcomplementary variables modeling
theory to biology and ecology. In:Problem Analysis in Science and Engineering. F. H. Branin, Jr. and K. Huseyin (Eds.). New York Academic Press,
pp. 427-462.

62

Jonesand Luyten

Swaney, D. P., J. W. Jones, W. G. Boggess, G. G. Wilkerson, and J. W. Mishoe,

1983. A crop simulation model for evaluation of within season irrigation
decisions. Trans, A S A E 26(2):362-568.
Thornley J. H. M., and I. R. Johnson. 1990. Plant and Crop Modezing. New
York Oxford Univ, Press.
Thornton, P. K., G. Hoogenboom, P.W. Wilkens, and W.T. Bowen. 1995. A
computer program to analyze multi-season crop model outputs. Agron. J.
87131-136.
van Kraalingen, D. W.G. 1995. The FSE System for Crop Sim~Zation,Version
2.1. ~ ~ a n t i t u tApproaches
i~e
in Systems Analysis, No. July 1995. Wageningen, The Netherlands: DL0 Research Institute for Agrobiology and Soil
Fertility and C. T. de Wit Graduate School for Production Ecology.
Visual Solutions. 1990. VisSim, Users M a n ~ a lWestford,
.
MA: VisualSolutions
Inc.
Wolf, S., J. Rudich, A. Marani, and Y. Rekah. 1986.Predicting harvesting date
of processing tomatoes by a simulation model. I; Am. Soc. Hort. Sci. ZZZ(1):
11-16.

Cornell University, Ithaca, New York

Crop modelings considerable popularity and success as a research

and planning tool stems from its focus on the behavior of agronomic
crops as a system. In addition to the inherent biological complexity
of plant growth and development, there are numerous environmental
conditions and insults such as drought and attacks by pests and diseases that must be considered in a truly comprehensive characterization of crop production processes and yield.
Two tools have come to occupy a central role in dealing with
this complexity: mathematics and computers. Thefirst provides a
powerful language and framework for the conceptua~zationof the
relationships, and the second provides a powerful technique for exploring the relationships encoded mathematically. This chapter explores the process of representing physical reality using mathematics
and the process of extracting insights from these representations. In
other words, I explore the process of using mathematics as a working
tool to gain insight into a problem.
63

This discussion consists of

l. A brief review of the process scientists use to conceptualize
the regularities of nature for predictive purposes
2. A brief discussion of the role of mathematical models in connecting the conceptual and experimental tasks
3. An extended discussion of the practical steps engineers have
found useful in formulating, validating, and interp~eti~g
mathematical models
4. A consideration of a representative problem to illustrate and
to make the discussion of the traditional modeling process
more concrete

.
The process of creating new knowledge involves the exploration of a
focus question or puzzlement. Figure l depicts this process (Novak
and Gowan, 1984). Guided by the focus question phenomena of interest are examined. Beginning at the base of Gowins Knowledge
Vee certain events or objects are observed. eaning is attached to
these events or objects by an intimate interplay of conceptual (or
t h i n ~ g and
) methodological (or doing) considerations, as depicted
by the two sides of the Vee.
Consider first the methodol~gicalconsiderations (on the right
side). Records of the events or objects are made. The facts are then
ordered or otherwise organized (transformations). The results are
then presented as data using tables, graphs, etc. This is followed by
interpretations, explanations and genera~ations. These lead to
knowledge claims about what was found. The final step consists of
value claims about the meaning or significance of the knowledge
claims.
On the conceptual side of the process, ~ e g i n n i n g
at the top-left
of the Tee); going from the most general, i.e., World View
(that
nature is orderly and knowable) to Philosophies, to Theories (logically related concepts) to Principles (derived from prior knowledge
claims) to Constructs and Conceptual Structures to Concepts (signifying regularities). During the creation of new knowledge there is an
active interplay between the conceptual and methodological sides of

nature is orderly and

knowable)

in field or out of field, of the

claims produced in an inquiry

generated by theory; FQs focus

attention on events and objects

nowledge Claims:

New
generalizations, in answer to the
telling questions, produced in the
context of inquiry according to

Toulmin)

reasoning leading to explanations

Inte~retations,Explanations,
~ e n e ~ i i z a t i o n sProduct
:
of

nnciples:

Conceptual rules governing

the linking
patterns
of
irt events; propositional in form; derived from prior
knowledge claims

methodologyprior
and
knowledge used
warrent of claims.

esults:

Representation of the data in

tables,
graphs
and
charts

constructs: Ideas which support reliable

theory, but without direct referents in events

ransformations:

or objects.

Ordered facts governed

by theory of measurement and classification

Conceptual Structures: Subsets of

directlyinused

theory

acts:judgment,
The

based on trust in method,

that records of events or objects are valid.

the inquiry

ecords of Events or Objects

concepts:

Signs or symbols signifying regularities

in events and shared socially

Events/Objects:
Phenomena of interest apprehended
through concepts and record-marking:
occurrences, objects

IGURE
Thevee heuristic with descriptions of the interacting elements
involved in the construction or analysis of knowledge in any discipline. Although all elements are involved in any coherent research program, the major sources of difficulty in individual inquiryusually begin at the bottom of
the vee, where concepts, events/objects, and records must be scrutinized.
(From Novak and Gowin, 1984, Reprinted with permission of Cambridge
University Press, New York, New York.)

the Vee. Novak and Gowin (1984) provide numerous specific illustrative examples from a range of subjects.
In a nutshell,this is the process we use to create new knowledge.
The process is so familiar that we often take it for granted.
being conscious of these steps makes the modeling process clearer; it
is fundamentally an artistic process in which there is no single right

CQQke

model.It is fundamentally a matter of successive approximations

deeply affected by matters of taste.
TOOLS

Scientific inquiry progresses more rapidly when there is a rich interplay between the conceptual and experimental at each step in the
process of creating knowledge rather than relying upon only one or
the other of these legs of the vee. However, the literature is replete
with disciplines in which the theoretical and experimental aspects
have matured at d~amaticallyd ~ e r e nrates.
t
Often, however, we as individual scientists rely more heavily
upon either one or the other, depending upon our training. For example, Fry (1941) contrasted the habits of thought of mathematicians
and engineers:
[l] Greaterconfidence in: Thetypical mathematician
feels greater confidence in a conclusion reached by careful
reasoning and is not convinced to the same degree by experimental evidence.
An engineer has less interest in pure forms which
have no apparent connection to the world of physical reality;
but may have great interest in such useful information as a
table of hardness which may be totally unrelated to any theory and which the typical mathematician would find quite
boring.
[Z] Severe criticismof details: A mathematician is highly
critical towards the details of a demonstration. For almostany
other class of people, an argument may be good enough, even
though some minor question remains open. For a mathematician an a r ~ m e nist either perfect in every detail, in form as
well as substance, or elseit is wrong. The mathematician calls
this rigorous thinking and it is necessary for work to have
permanent value; the engineer calls it hair-splitting and says
if indulged in would never get anything done.
[3] Idealization: Themathematician tends toidealize any
situation-gases are ideal, conductors, gases perfect, surfaces smooth. What the mathematician calls getting down
to essentials, the engineer is likely to dub somewhat contemptuously as ignoring the facts.
141 Generality: When confronted with solving

x3-1=0

Using ~ a t ~ e m a t i cas
s a roblem-Solving Tool

that simple question is likely to be turned into solving

Xn

0.

The mathematician calls this conserving energy. The

engineer is likely to regard this as wasting time.
One approach is not superior to the other. They are different and
complementary.
One of the truly great contributions that engineering can make
to crop modeling is the transfer of some of the concepts and techniques that have proven to be so fruitful for engineering purposes.
Far greater reliance has been placed upon the use of mathematical
techniques in the physical sciences than has been the case in the biological and agricultural sciences. Nevertheless, the use of empirical
models has gained widespread acceptance in crop modeling in the
past 30 years.
While a major portion of the body of knowledge associated with
engineering in general and in crop modeling in particular is heavily
empirical in nature, fairly extensive use has nevertheless been made
of concepts and theories that can be most conveniently expressed in
mathematical form.Thefollowing
discussion applies equally for
modeling, which relies almost exclusively on the use of computers
and empirical relationships, and the more traditional approach,
which relies on computers and first principles. The latter approach,
when enough is known to make the approach feasible, often provides
deeper insight and usually canbemore readily generalized to a
broader context.
. T

ELS

At the interface region between engineering and biology there persists the lingering and mistaken belief that biologically based problems are inherently so complex that they simply are not amenable to
mathematical treatment. One might imagine that this same attitude
could have been applied to quantum mechanics, were it not for the
astonishing success achieved using mathematics as a central tool.
Experimentalists, especiallythose who ordinarily work withbiological problems, are sometimes inclined to be suspicious of mathematical work. Although a healthy skepticism is highly desirable, the
role of mathematical modeling has many times been unfairly criticized (but certainly not in all cases). A particularly common difficulty
is the failure to note the necessity and desirability of making assump-

tions-in both mathematical and experimental studies. In both instances attention is directed to a limited problem and theconclusions
reached are valid only within the constraints of "the experimental
assumptions." ath he ma tical models (Noble, 1967) can be used to
cause attention to be directed to the most hportant questions in an
experimental study. If data agree with a theory, the level of confidence
in the predictive power of the model is enhanced. In other instances,
mathematics can be used to reduce or eliminate certain experimentation.
Nahikian (1964) graphically depicted the successive approximations associated with a mathematical model (see Fig.2). ~ e ~ n n ~
with a clear statement of a "problem," the most important properties
of a physical or biological problem to be studied are abstracted and
isolated by making judicious use of assumptions (left side of Pig. 2).
A mathematical problem, using basic physicalor chemical principles,
is developed to embody the "essence" of the problem. The "mathematical crank" (bottom) is turned and checked. The implications of
the model are articulated. These implications are then related to the
original problem through the observed phenomena. Discrepancies are
noted, and the cycle begins again. However, notice that the arrows
are bidirectional throughout, indicating that the logical progression
may reverse at any point.
The advent of the computer is having a great liberating effect
on the use of mathematics. There is no longer the overriding need to
formulate problems to conform to the limited tools of classical anal-

Define problem, Make

s i ~ p ~ i f y i nassumptions,
g
Review basic principles

Make mathematical checks,

Review assumptions, Relate
model to original problem

Development of mathematical models. [Adapted from Nahikian

.]

r o ~ ~ e ~ - S o lTool
vin~

ysis. Onthe other hand, when using a computer we still must abstract
and idealize a given engineering problem. We still must break the
original problem into its elements, decide what is significant, specify
the data we must start from, what the steps of the analysis are, and
what end results are required. These steps require the habit of
thought of the mathematician.

The process of forming a model (left side of Fig. 2) is both a science

and an art form. The following discussion concentrates on the wellestablished techniques used in creating and validating a model and
gaining insight from it. The more artistic, but just as essential, aspect
of modeling is mentioned only in passing; nonetheless, that can be
the most important and decisive step. It is mainly here that a reasonable balance is struck between complexity and simplicity. A model
can be so inclusive of details that it becomes intractable or difficult
to understand. A model might be endowed with so many empirical
coefficients, discernible only from experimental measurements for a
particular setting, that it has limited utility in a setting that is only
slightly different because of a loss of predictive power and insight
into the underlying mechanisms. On the other hand, a model can be
made so simple that the essence of the situation is lost. For example,
certain phenomena might not be encompassed if represented using a
system of linear ordinary differential equations rather than a system
of nonlinear ordinary differential equations. A linear model, for example, does not encompass stable limit cycle behavior. And a deterministic model might be unsuited for situations in which probabilistic
models might be successful.
Suppose you are interested in examining the temperature variations within the walls of an internal combustion engine during the
combustion cycle. How might you represent the problem as a mathematically tractable problem, yet achieve a good approximation to
experimental values?
You probably would formulate this problem as one of heat conduction in a solid. Next you would probably realize that you must
define an appropriate geometric c o n ~ ~ r a t i oDealing
n.
with the actual geometry would probably require that you resort to a numerical
approach such as the use of finite-difference or finite-element techniques. Suppose you wish to make a simpler first approximation using solutions from the existing literature?

Cooke

Several possibilities cometo mind. You might formulate this as

(a) a thick cylinder subjected to a periodic temperature fluctuation on
the inner wall of the cylinder or (b) a cylindrical hole in an infinite
space. Case a will lead to the use of Hankel functions, whereas case
b willlead to the use of the simpler Bessel functions.Depending upon
the frequency of the periodic fluctuation at the inner cylinder, the
depth of penetration might be approximately the same in both cases.
If so, the simpler model might be adequate.
On the other hand, if the curvature of the inner cylinder is not
a ~~a~
consideration, you might fornulate this as (c) a finitethickness slab or (d) a semi-infinite solid (a half-space having a flat
edge). For casec, the solution will involve hyperbolic functions, while
case d will involve the even simpler trigonometric functions (or the
complex exponential function). As it turns out, the ratio of periodic
surface temperature to the periodic temperature at any depth in the
wall is roughly the same for all four cases when the frequency is in
the range typical of an engine, but the mathematical burden becomes
very much less for the simpler case d.
A crop modeler probably would recognize an analogy with the
daily or annual periodic temperature fluctuations in the soil and how

FIGURE3 Alternative geometries. (a)A thick cylinder subjected to a periodic

temperature fluctuation on the inner wall of the cylinder. (b) A cylindrical
hole in an infinite space. (c) A finite-thicknessslab. (d) A semi-finite solid.

Using ~ a t ~ e m a t i c s a ro~lem-Solving

71

the temperature varies with distance into the soil. Recognizing analogies can greatly strengthen ones confidence in a solution.
The greater the insight into the underlying processes you can
muster at any stage of development, the more likely an acceptable
balance between simplicity and complexity canbe achieved. The crucial point to remember is that you are using a mathematical model
to gain insight and that no single representation or model serves all
purposes. A novice modeler is tempted to beg for a universal representation that fits all situations, but alas, if that is your goal, you are
likely to trade an obscuring complexity for the clarity that often can
be achieved more readily through simplicity.
Bowen (1966) described the basic issues to be considered in formulating a problem (Fig. 4). Before a problem can be properly defined, the decision maker must be identified. Usually the decision
maker is both the person or group whose objectives are not being
met and the one who controls the resources needed to effect a solution. Next the decision makers objective and resources must be clarified and weighted. Usually objectives conflict and cannot be simultaneously satisfied. Furthermore, the environment with its associated
constraints must be specifiedand understood. A solution t ~ignores
~ t

Resources

Selection of Decision
Maker

I
1

ProblemDefine

Test solution or machine

I

Acceptance or rejection of solution

FIGURE4 Steps in formulating

problem.

these const~aintsis not a solution, Only after all three of these issues

(objectives, resources, and environment) are understood clearly can

the problem be properly posed or defined with sufficient precision
and clarity. How you define the problem fundamentally shapes the
rest of the process!
To illustrate how a model fundamentally depends uponhow the
problem is framed, consider the approaches taken in solving the problem of mechanically harvesting tree fruits. The problem has been
manifest in the literature in at least two different forms. An earlier
view held that the criterion for fruit removal was the tensile forces
developed from inertial reaction of the fruit stem while the tree is
being vibrated. This inevitably led researchers to center on ways to
produce vigorous shaking of the tree, e.g., at one of the natural frequencies of its limbs. On the other hand, if one defines the problem
in terms of the bending stresses in the stem, at the fruit or at the
branch, then the argument follows a substantially different pattern.
The fruit and stem assembly can be regarded as a double physical
pendulum, and one might then study the frequencies that produce
the largest motions of the fruit-stem system rather than of the supporting tree structure. One should not be surprised to learn that the
inferences concerningharvesting are quite different forthe two statements of the problem and result in fundamentally different levels of
understanding.
Only after a clear definition has been achieved-and this is often the most crucialand difficult phase in solving a problem-should
you begin the process of generating alternative solutions. This stage
requires imagination as well as a capacity to tap all your previous
professional experienceand knowledge. Novel associationsand analogies often play an important role. In short, creativity plays the crucial role in this step. Many techniques can be used to stimulate this
process. Often a mental block must be overcome if the problem is to
be recast from a fresh point of view. There is a powerful tendency to
view the problem in ways that are familiar even when objectives are
no longer being satisfied.
The selectionof the most logicalalternative usually requires that
a series of very specific problems be formulated and analyzed. This
process is explored next.
As you develop plans to implement the alternative selected, you
simultaneously articulate a quantifiable,or at least u n a m b i ~ o u s /
measure of performance by which you will judge the proposed solution. You do this before your plan is implemented and before you
must make a rational assessment of whether you have succeeded.

73

After your solution has been implemented, there is a responsibility to determine whether your solution works. This usually involves some form of experimentation and an evaluation of whether
you have met the objectives previously articulated.

Although numerous texts on mathematical modeling have been published, a very old text by Ver Planck and Teare (1954) serves well as
the focal point for this discussion. Table l summarizes the approach
presented in their book, but Ive also made some adaptation. These
are the time-honored steps that engineers routinely use in problem
solving. Because the steps are so familiar, their significance is often
overlooked.
Explicitly identifying these steps makes the process more comprehensible and more than just an art form. The steps are (A) define
the problem, (B) plan its treatment, (C) execute the steps, (D) check
your work, and (E) learn and generalize from your analysis. Lets
consider each of the steps in this process.

In dealing with non-textbook problems, great importance must be

placed upon undertaking a heuristic first consideration of the problem before delving into the intricacies of a particular approach. In
other words, consciously and explicitly decide among alternative
approaches before you commit your time to any specific one.
By stating the anticipated results quite early in the study of a
problem, you usually should be better able to judge the validity of
your analysis. This will provide a basis for guiding the analysis as
well as a background against which the analysis can be appraised.
The appropriate point at which to review the relevant literature may
well depend upon the natureof the problem. To avoid reinventing
the wheel, the literature might come very early but on the other
hand there is the possibility of prematurely channeling ones thinking to current ideas. There is also a need here for moderation, balance, and judgment.
How you state the problem to be solved, i.e., frame the puzzlement, will have a profound impact on the entire process. If you
solve the wrong problem, your time may be wasted. If the problem

Table 1 Aspects of Using Mathematics in Engineering Problem Solving

A. Define the problem.
1. Divide the problem into a series of specific questions.
2. Discuss the problem in a qualitative fashion and present a heuristic
motivation before plunging into equations.
Consciously decide among alternative approaches.
4. State what results you anticipate. This is equivalent to a hypothesis.
5.Review the relevant literature.
B. Plan its treatment.
1. Review the relevant physical or chemical principles. Express them verbally
and then mathematic all^.
2. Identify the assumptions that may be desirable and/or necessary.
3. Formulate the problemmathematically.
a. Usesketchesfreely.
b. Define all symbols and give the associated units.
c. Show all coordinates with positive directions identified.
d. Identify all assumptions as they are embedded in the model.
e.Citeallreferences
consulted.
C.Execute the plan.
1. Use more than one approach when possible.
2 . When carrying out the mathematical steps, number all equations.
3. Use frequent intermediate checks.
4. Provide an account of the motivation for each step, and indicate the logical
connection between steps.
5. If an impasse is reached, simplify the problem by altering the assumptions
and/ or the problem definition.
D.Checkthoroughly.
1. Check dimensions and units.
2. Check limiting cases.
3.Checksymmetry.
4. Checkreasonableness.
5. Check range of variables.
6. Reexamine all assumptions for probable effect and importance.
7.Use alternative derivation.
8. Useanalogies.
E. Learn and generalize from your analysis.
1. Summarize the important findings, including limitations.
2. Interpret the equations in terms of the original problem.
3. Have important factors been overlooked?
4. What is important by its absence?
5. Use dimensionless numbers in the graphical presentation.
6. Verbally state the meaning of the final mathematical equations.
7. Can the proposed scheme be improved or altered so that it will not work?
8. Checknumericalcasesforclarity
and feasibility.This is a check of the
magnitude of the effect and the required parametervalues.
9. How much parameter variation can be tolerated?
10. How many significant figures should be used?
11. Is the graphical presentation fully clarified, including clearly labeled axes,
and is an interpretation of the information presented?
12. What uncertainties still exist?
13. How does the model relate to the original problem?
14. Can and should the problem now be simplified or generalized?

Source: Adapted from Ver Planck and Teare (1954).

r o b l e ~ - S o l v Tool
i~~

75

is stated carelessly, you may needlessly complicate your task. The

clearer your grasp of the problem, the easier the taskof defining the
problem becomes. A poorly defined problem will likely increase the
complexity of the mathematical formulation.

A verbal discussion of the problem (recorded in your notebook) is

frequently important in refreshing ones grasp of the applicable physical principles, because the mathematical statement of the physical
principles, especiallyif presented in mathematical form, maybe terse
and will usually contain many assumptions that are not immediately
obvious. The actual formulation of the mathematical model should
be accompanied by a very careful statement and review of the working assumptions that you must make in order to form a tractable
mathematical model. Re continually aware of the trade-off between
simplicity and complexity. Quite often it is advisable to formulate and
solve a simplistic model before adding the level of detail that you
believe will eventually be desirable. Whenever feasible, use multiple
approaches. If the same result can be obtained by alternative methods, e.g., conservation of energy, and through equations of motion,
you will increase your confidence in the results.
Furthermore, the validity and usefulness of your analysis usually suffer far more froman important unacknowledged assumption
than from important acknowledged assumptions. Remember that
modeling is an iterative process, i.e., not every detail is or should be
included at the outset. In the spirit of Occams razor, a successful
simpler explanation is preferred to a successful but complex explanation. Formulating and solving one or more highly simplified models as a first approximation is usually prudent butis often overlooked.
Often a complex problem can be solved as a sequence of independent subproblems that, considered together, describe the more
complex situation. Often several simpler problems will allow you to
bracket the solution sufficiently to make a more sophisticated analysis
unnecessary.

When you first introduce a variable, immediately define it in wor

and give its associated units. This will facilitate comprehension and
subsequent checking. All equations must be dimensionally homogeneous,i.e., when the units replace the variables
e equation, the
units on both sides of the equation must agre
wary of using

numerical constants that have associated units; it is better to assign

a parameter name. The parameter name is then treated on the same
basis as thevariables, and the dimensional homogeneity checkis less
error-prone.
Use sketches freely, number them, and provide captions. Even
a crude sketch can often provide the clarity that would otherwise
require many words. Label the axes, the positive directions, and all
the components of the figure. Captions will be useful when you return toyour work after any substantial time lapse. Likewise, for later
recall and for use in the formal report writing stemming from your
analysis, you should cite all the references you consulted. A methodical approach encourages careful thinking.
To facilitate the derivation and discussion of the equations, number every equation, even though you may choose to leave some unnumbered in your formal report. Remember that your analysis will
progress more smoothly if you capture your thought as it proceeds.
This not only helps you think through the problem, but also is an
invaluable aid when you return to your analysis later. As the process
proceeds, record your thoughts about the motivation of the analysis.
Always provide a verbal account of the logical connections and the
details in the analysis. You will not be well-served by dismissing the
details with an assertion that "it can be shown.'' Also, your notes
serve an archival purpose but should be treated as distinct from a
formal report, which is used to present your ideas in an economical
form for the benefit of others. Your notes are for your own use.
If you reach an impasse in theanalysis, simplifythe problem by
making additional assumptions or by modifying the problem definition. Once you understand a simplified version, even if it is too
simple to be included in the final report, it may provide just the
insight you required to solve the more complicated version.

Thissection draws heavily upon two primary,out-of print

sources: Ver Planck and Teare (1954, chapter
and Johnson (1944,
chapter
"Engineering work not thoroughly checked has little value."
(Ver Planck and Teare,1954, p. 229). Frequent use of intermediate
checks is highly desirable. Your analysis must be free from essential
errors-from using the incorrect methodology to getting the decimal
point wrong. Your professional reputation depends upon your reliability so you must devote serious attention to checking and validating

77

your work. In a college course perhaps the greatest harm resulting

from your errors is to your course grade, but inprofessional practice,
the safety and health of other persons may beat risk. Incontrast with
the classroom situation, reliability is vastly more important than
speed. Also, when you publish a paper, an error remains for all the
world to see and perhaps your errors may mislead others. Because
most engineering problems are solved over a period of days rather
than minutes, it is important to develop the habit of providing the
appropriate step-by-step documentation and checking each step in
your analysis. Engineering projects often involve team effort, but individuals retain responsibility for assuring the accuracy of their work.
Team confidence in you depends fundamentally upon your ability to
produce carefully checked work.
By far the most effective way to check an analysis is by means
of a properly designed experiment. (Ver Planck and Teare, 1954, p.
229). Depending upon ones perspective and experience, as the earlier
Fry quote suggests, one will be more easily convincedof the validity
of an analysis by experiment or by careful reasoning. The scientific
method holds that a properly performed experiment serves as the
final court of appeal when analysis and experiment differ. However,
this does not mean that you must immediately and uniformly resort
to expe~mentation-neglecting your mathematical tools. In the first
place, experiments often require that you perform some nontrivial
analysis to properly construct the experiment. Experiments are often
expensive and time-consum~g.Furthermore, the task may be inherently more amenable to either experimental or mathematical modeling. Your relative level of expertise as a modeler or as an experimentalist, independent of the particular project, may influence which
path
is less costly to pursue.
ofessional method implies not only checking an end
reasonable way but also checking frequently as the
work progresses so that the analysis is accurate at every sta e of its
development. (Ver Planck and Teare, 1954, p. 229-230). C
so important that it should become instinctive and hab
doubtedly much of your checking
S willoccurafter
you have
esults that look promising.
ver, checking must
at every stage of your ana
ot just after the de
rrors, even simple and easily detected ones,
your analysis and necessitate lost time retracing your steps.
In this chapter we are specifically interested in techniques that
make your use of mathematical tools more effective (for validation

Cooke

and to help you gain insight), so lets consider some specific techniques, including some that can beapplied rather easily as your work
progresses.
Kinds

Checking

Two other kinds of checking are especially useful: (l)overall checks.

The latter often mergewith learning and generalizing, and (2) checks
of the manipulations (arithmetic, dimensional, etc.).
~ ~ e c ~ by
i nxperience
g
omp paring your resuEts with your previous e ~ ~ e r i e n c eand
s common sense constitute the most i m ~ o r t a n t but
, often
overzoo~ed form
of c ~ e c ~ i n If
g . a calculationproduces a result that is off
by an order of magnitude (e.g., a car weighs twenty tons), your internal alarm should warn you. t i s you develop more experience in a
particular specialty, your sense of scale becomes a powerful asset.
Sometimes conversion of numbers into a more familiar set of units
canfacilitate an order of magnitude check with your experience.
Checking by experience is extremelyvaluable, but of course, not
infallible.
i ~ e n s i ~ ~ ~ ~
Toerepresent
c ~ i n ga physically meaningful relationship, an equation must be dimensionally homogeneous. For example, each term in an equation consisting of a sum must have the
same units. This is a necessary, but not a sufficient condition. If the
units are not consistent, an equation cannot be valid, but a dimensionally homogeneous equation may still be incorrect. That is,failure
to meet this test is positive proof of incorrectness, but meeting the
test does not establish its correctness. For example,
missing terms will
not be detected, nor can you be sure that dimensionless parameters
have the correct magnitude or sign.

Names and Units

Ver Planck and Teare (1954, p. 231) illustrate the need to use the
same units with the following:
An equation is written with the symbol for each quantity followed by the name of the units in which that quantity is
expressed.
5 apples

apples = 11 apples

is dimensionally homogeneous, but

5 apples
is nonsense.

+ 1 cow =

apples

Using ath he ma tics as a ~ r o ~ ~ e ~ - S o lTool

ving

79

A more representative example:

A (baskets)

The units for each term are treated algebraically. Contemporary symbolic mathematical software can be used to automate
this task, as well as handle the mundane but error-prone task
of conversion of units.
There are a few special situations to be considered. Numerical constants in an equation may represent pure numbers
(e.g.,
but may also be dependent upon a particular set of
units. For this reason, symbolic parameters are usually preferred because their units canbechecked
more easily if
treated on a basis symmetrical with the variables. For example, if the variable in the quadratic equation has the units
of meters, then clearly the parameters b, and c must have
units too.
The arguments of certain special functions, such as the
trigonometric functions, exponential functions, and Bessel
functions, are pure numbers and must be dimensionless. Beware that using units of measure such as degrees rather than
radians can createan apparent discrepancy. Similarly because
the logarithm of a product can be separated, an apparent discrepancy may exist. Sometimesa physical law must be introduced to complete the conversion process to demonstrate dimensional homogeneity.

This is similar to units check, but each factor in the equaiton is replaced by its dimensional formula in terms of primary quantities such
as mass (M), length (L), and time (T). Unlike the units check, this
does not verify consistency of system units. Ver Planck and Teare (p.
illustrate the procedure:

. . . density is mass per unit volume,or mass divided by

length cubed, hence [M Ld3] where M represents the dimensions of mass and L the dimension of length. Similarly, it
might be desirable to have a dimensional formula for electrical resistance in terms of the dimensions voltage [V], charge

, and time [TI; it is [VG-IT], as may be deduced from

ms Law and the definition of current as a charge passing
through a circuit per unit time. . , If too many primary quantities happen to be chosen in a particular case, it will be necessary to e~minateone or more of them by using known
physical relationships, For instance, in a problem in dynamics
if all four of the dimensions, force [F], mass [M], length [L],
and time [TI are used as primary ~uantities,it will be found
necessary to e ~ m ~ aone
t e of them, This is done very easily
by using Newtons law f = ma; it shows that force [F] has the
dimensional formula [MLT-2], or,if you prefer, that mass [M]
has the formula [FL-*T2].
Case Checks

Limiting case checks are a particularly useful technique. Often your

erience willsuggest whether an expression should increase ordecrease as each parameter is increased or decreased. Should the result
increase or decrease without bound as a parameter or variable is
changed, or should there be an asymptotic bound? Or should there
be a relative maximum or minimum?
Checking a special value, such as 0, 1, e, multiples of IT,or infinity is often especially easy and revealing, Checking these special
values often permit you to compare the expression with your intuitive sense of what should happen.
Identifying a steady state behavior often can be identified. An
independent derivation of a simpler limiting case may be possible;
such indepe~dentconfirmation can greatly strengthen your confidence in the more general derivation.
Ver Planck and Teare (p. 46) illustrate this technique using a
simple exponential result from heat transfer. The temperature T approaches the initial value T = 0 as time approaches zero and T aproaches the steady state
as time increases without bound.

Ast40,

Limiting case checks also plays an important role in helping you

grasp the underlying physical significanceof your made1 and is often
useful in the m o d e ~ n g
process when you are interpreting the
significance of your result.

Symmetry as a Check

Symmetry also provides a powerful means for checking your derivations. Checking a result for even and odd functions is easy and
often helpful. For example, if a variable has a positive direction, what
result is expected if the sign of that variable is reversed? Will the sign
of the expression be the same or reversed?

Ver Planck and Teare (p. 242) provide two examples of using symmetry to identify potential errors.

T =hA
L [~1 - exp

(-~)]

Ast-0,

T-0

Ast-+a,

T+- 4

hA

(Ver Planck and Teare, 1954, p.

The equivalent resistance for the following circuit (see

diagram) is known to be

Suppose your equation were

IGURE

Circuit diagram.

82

Cooke

Because R, and R, are similarly placed, they of course, should

be similarly placed in the equation (Ver Planck and Teare,
1954, p. 243).
Notice that the error below could not be detected this
way.

Consider the following spring.

Suppose the relationship derived for the deflection of a
leaf spring were

where
P = load (lb)
= deflection at the load (in.)
E = Young's modulus for the leaf material (lb in.-')
b = width of leaves (in.)
n = number of leaves
h = thickness of each leaf (in.)
Zl, E,E, = lengths (in.) as shown

E, b, n, and h should be in the numerator, for an increase

would tend to increase the stiffness of the spring (Ver Planck
and Teare, 1954, p. 244).
Placement of E's is less clear. Having E, in the denominator is less clear. ButE, and E, should enter the formula symmetrically.Discover that subscripts 1 and 3 were swapped
(Ver Planck and Teare, 1954, p. 244).

I
I

11

FIGURE6 Spring schematic.

Using ~ a t h e ~ ~ tas
i cas~ro~lem-Solviflg
Tool

Let f = / and 1 - f = /
would read

Then the corrected formula

Here f and (1 - f ) enter symmetrically (Ver Planck and Teare,

1954, p. 245).
~ e ~ e r afrQm
~ i ~Y e
Q~r-

An analysis should always be concluded with a summary of important findings, with particular emphasis on the limitations imposed by
the assumptions. Of course, all important assumptions should be
made explicit rather than implicit. Mathematical equations are not
useful to the engineer until they have been interpreted in terms of
the original problem. You must verbally state the meaning of the
appropriate mathematical equations. If you cannot state their meaning, quite lilcely you do not yet understand them.
The analysis should usually but not always, involve numerical
testing of the results. This will clarify the meaning and the feasibility
of the particular problem. Thisprovides a sense of scale, alerting you
to gross blunders or possibly to your lack of success in finding a
solution. Finally the ident~ableuncertainties that still exist should
be clearly stated. Mistakes follow if you carelessly exceed the conditions on which your analysis is based. The analysis should not be
concluded until the possibility of simplifymg and/or generalizing the
analysis has been considered, explicitly and individually.
Consider now a specific analysis in order to make the foregoing
discussion more concrete. This alsoillustrates how a simple analysis
can make an expensive experiment unnecessary.

In an effort to produce a morerational approach to thedesign of field

machines and operations affecting the planting of a ~ c u l t u r acrops,
l
various factors must be investigated. Soil temperature is known to be
one of the important factors affecting seed germination. It is also
known that once germination has been initiated, the rate of internal
heat generation due to respiration increases rapidly.Can this thermal
effect be measured in situ with thermocouples?

Does the heat of respiration of a germinating seed appreciably alter

the soil temperature and internal seed temperature? And can this be
measured in situ?
If the other factors affecting germination are not limiting,
e.g.,available soil moisture, suitable soil temperature,
adequate oxygen supply acceptable soil physical impedance, no inhibiting chemicals (naturaland artificial),
previous temperature history, and activation of phytochrome by proper wavelength of light, the seed is viable.
The heat transfer due to external sources (e.g., diurnal
fluctuations) and the heat of vaporization and heat of
wetting are neglected in this first analysis.
The soil is homogeneous and isotropic with respect to
thermal properties.
Only the steady-state magnitudes are to be explored (at
first). In other words, the seed and soil will be assumed
to be in a (dynamic) thermal equilibrium independent of
time, with the temperature never near freezing (to avoid
latent heat difficulties).
The thermal constants are independent of time, temperature, and position (for seed and for soil).
The seed is taken to be (initially) spherical of radius Y =
a (which would be valid for certain seeds), to be producing heat internally at the uniform, constant rate A. per
unit volume per unit time, and to have thermal conductivity KO.
No heat is produced in the soil, which has conductivity
K.
There is imperfect thermal contact between the seed and
soil such that there is a film temperature drop at the seed
surface. Specificallythe thermal energy conducted to the
seed surface passes into the soil at a rate proportional to
the difference tem~eraturebetween the seed and soil.
The seed is sufficiently far below the surface of the soil
as to be relatively
unaffected
by the soil-to-air

disc on ti nu it^

eat transfer within the seed and within the soil is by

conduction only i.e., heat transfer due to moisture migration will be assumed negligible.

at~ematicsas

~ r o ~ l e ~ - S Q lTool
vi~g

FIGURE7 Seed-soilschematic.

These assumptions are shown schematically in Fig. 7.

Since the soil temperature at large distances from the center of the
seed will not be affected appreciably by the heat generated by the
seed within the soil (which would be at uniform temperature if the
seed were not present), that temperature may betaken as a reference.
In other words, all other temperatures may be calculated with
lim vu2= 0
r-m

as the zero reference.

From Assumption 10, we will use the heat conduction equation,
with heat generation, as given by Carslaw and Jaeger (1959, p. 10).

where
V

= scaled

temperature, C" (cgs units)

= diffusivity, cm2/s
= heat generation rate, cal/ (S cm3)
K = thermal conductivity, cal/ [S cm2(OC/ cm)]

and
K

where is density (bulk) (g/cm3) and c is specific heat (at constant

pressure approximately).
From Assumption 4, we may set

In spherical polar coordinates (due to AS),

Cooke

[1?1a (P

"> -1

d2v
d e sin2@
Due to the symmetry in 0 and
Assumptions 6 and 7, Eq. (1)may
be applied to the seed and soil separately:
~v

= --p

sin 0 80 (sin 0

+,

[?,-(..~)]

"p
l a

0,

Y >

The conditions required for solution of Eqs. (4) and (5) are as
follows. The temperature at great distances from the seed must approach the reference value of zero.
lim v2(r) = 0
(6)
where v, is not a temperature but a temperature difference-Celsius
degrees from the reference temperature.
In order for the boundary between the seed and soil not to increase in temperature without bound, the continuity relation must be
imposed. The flowof heat per unit area per unit time in the direction
of increasing variable is given by
= "v
(7)
where
dv a, dv
a+ d v
-+-"l""
ar Y 80 Y sin 0 a+
By symmetry, the temperature does not change in the 0 and directions for a surface Y = so
av
"- a+ = O
At the interface (Y =

so

The minus sign appears in Eqs. (7) and (11)because the flow
occurs in the direction of decreasing temperature, i.e., d v / a r will contain an inherent negative sign td compensate for the minus sign in
Eq.

Using ~ a t ~ e m a t i cass a ~ r o ~ ~ e m - s o ~Tool

ving

From Assumption 8, we have, in place of a requirement for the

continuity of temperature, the following:
(12)

where H > 0 is a proportionality constant that is called the coefficient

of surface heat transfer (Carslaw and Jaeger, 1959, p. 19). (The H of
Carslaw and Jaeger corresponds to the h more widely used in the
literature on convective processes.) The
surface thermal resistance per
unit area may also be defined as

XZE-1
H
This H should not be confused with the one in the expression
H
h=K
Equations (4) and (5) may be solved by integration and using
the boundary conditions (6), (ll), and (12).
Carslaw and Jaeger (1959, p. 232) give (without proof) the following solution:

+ 2XaK0 + 2a2(K0/K)]/6K0,

= Ao[a2 = A0a3/3Kr,

a(15)

>a

(16)

Equations (15) and (16) satisfy the differential equations (4) and
(5) and the associated boundary conditions.
Since solutions to Poisson's equation are known to be unique,
we may be confident that the actual solution has been found.

Now that the temperature has been found at all points, a check
should be made.
~~~~~

~ Suppose
e
c heat
~ generation were zero. Then

lim

=0

OK

lim

=0

OK

Ao-0

Suppose the size of the seed becomes vanishingly small. Then

lim
a-0

=0

OK for

l i m vO, K
=O
forr>a

If R

0, then continuity in temperature at the interface should exist,

lim[lim v,] = lim v,

R40

v"a-

OK

rea+

If the thermal conductivity of the soil is very great (i.e., a very

conductor), then the heat would be conducted away and the
soil temperat.ure should not be affected, since the mass of the soil is
assumed to be infinite. The seed temperature should be finite.
lim v2 = 0
K"&

lim v, = Ao(a2- r2 + 2RaK0)/6K0 = finite

K40

On the other hand, if the soil behaves as a perfect insulator,then

the soil temperature would become very great for finite r, and the
seed temperature would increase without bound.
lim v, = 0
0, all finite r
X-0

1im v3 =

all r c a

K-0

Other limiting checks may be performed, but a sense of underand confidence has begun to develop.
S embed this problem in a more general context to clarify
what has been learned.
In order to obtain the results in terms of the rate of heat prouction per seed, the flux over the surface Y = a may be obtained
by integration.

= (4/3)n;a3A0

Using equation.

and equation (29), then

(29)

v, =

Q
~

4TKa

This is the same temperature that would have resulted from an isolated point source in a medium of conductivity K at a distance r = a.

~
~~~e~~ It~ will nowebe instructive
~
to check
~ the results
~
numerically to establish reasonablebounds on the temperatures to be
expected. If we now assume intimate seed-soil thermal communication ( R = 0), i.e., seed coat temperature equals adjacent soil temperature, then the tempera~reof the center of the seed is found to be
A o a 2 &a2

v, = -+ -,
seed center
4Ko
3K
The seed surface temperature is
A,a
v, = 3K

seed surface

(33)

Then,

The soiltemperature at the surface may be obtained from either

Eq. (33) or Eq. (27). R may assume any value in Eq. (277, so we shall
use that result. (The value of R will affect the seed temperature but
not the soil temperature.)

v, = soil temperature at Y = C
Q = rate of heat production, cal/ (S. seed)
a = radius of seed, cm
K = soil thermal conductivity, cal/ [S cm2(C/ cm)]
~ ~ e r ~ a l C o n ~ ~ c Carslaw
t i v i ~ y and
:
Jaeger givefor average

soil
K = 0.0023 call [S cm2(oC/cm)]
Kernel Size:
A 12/44 in. round hole screen is used by the US.
Department of Agriculture (USDA) as a basis for determining corn
breakage. If we take twice that size, say radius of kernel
a =

(g (in>
in.)

2.54 cm = 0.475 cm

Heat ~ r o ~ ~ c oft iKernel

o ~ (Corn): From Prat (1952, p. 391) the
graphical and tabular data on respiration rates (cal/h) vs. time are

Cooke

90

given. Varieties of corn with peak respiration rates of 3.0,2.1, and 2.3
cal/ (h. g) with four seeds were given.
= [3.0 cal/ (h-g)]

(3ioi (-

0.000208 cal/(s.seed)

From Eq. (31),

0.000208 cal/ (S seed)
(12.56)(0.475 cm~(0.0023cal/ [S- c m 2 ~ C / c m ) ~ )
(35)
= 0,0151"C
The units check.
However, this is far below the level that can be measured with
athermocouple!! Perhaps a series arrangement of thermocouples
should be considered.

v, =

The experimental arrangement proposed will

probably not succeed.
~~~~s The temperature rise at the center of the seed
can also be computed from Eq. (34).
Let .R = 0. K for corn can be taken as 1.22 Btu.in./(ft2.hr."F)
(ASAE, 1967, p. 285). This must be converted to cgs units, where the
factor 0.00413 was obtained from Carslaw and Jaeger (1959, p. 3).

(=)

l .22 Btu.in*
(0,00413) = 0.00042
ft?* hr*"F 12 in.

Since

A,a
3K

v, = -

K
+ V, = 0.056"C
Zr;c,
This is the maximum temperature rise above the undisturbed soil
temperature. The temperature difference v, - v, = 0.041"C is also below the limits for a thermocouple. Under the most careful laboratory
conditions,a thermocouple can be used to detect differences no
smaller than (0.05OF) (5/9) = 0.03"C.

v, = -v, + v, =

In view of thermal conduction along a thermocouple and the difficulty of correctly placing the probe in the seed, this appears to be a
futile experimental approach to the problem. Thegradients that occur

Using ~ a t ~ ~ ~ aast ai c s

normally in the soil will be vastly greater than that produced by the
heat of respiration.
However, if the contact resistanceof the surface were great, then
the internal temperature might be affected enough to alter the biochemical reactions within the seed. Therefore, the problem may be
redefined, if desired, to explore this new problem suggested by the
analysis.
Seed and soil temperatures will frequently be below those of the
calorimeter tests with a consequently lower respiration rate. The water absorbed by the seed (imbibition) will,of course, alter the thermal
energy status of the seed. The thermal conductivity of soil changes
radically with moisture content. Also, the corn kernel is not spherical.
However, the above computations may be taken as bounds on the
actual values.
If the seed were very near the surface, the method of superposition of heat source and sink could be used. The other extreme case
of a solid spherical seed producing heat at the constant rate
per
unit time per unit volume in air at constant temperature can be
examined.
v = -[h(a2-

6hK

24,

25

<a

(37)

[See Carslaw and Jaeger (1959, p. 232).]

Finally, the heat of respiration is probably not produced uniforrnly throughout the seed. Specifically the embryo temperature
would be expected to be higher than that of the endosperm. This
could have important consequences for behavior of the seed but is
not likely to alter the soil temperature appreciably.
Heat of respiration experiments must necessarily be conducted
in microcalo~meters as
is the current practice of plant scientists. The
air temperature change within an adiabatic finite chamber due to
respiring seeds might be usefully examined.

The process of applying mathematics to the solution of engineering

problems is an old art form that is enjoying increased attention at the
interface between engineering and biology. Thenow ubiquitous presence of computational power unimagined only a few decades ago
has massively expanded the importance of and scope of interest in
mathematical tools and modeling. This chapter discusses the process
we use to connect this form of conceptual t h i n ~ n gto problem solv-

ing. ~lucidatingthe process-even those aspects that are obvious to

the practitioner-makes the process more usable and comfortable.
~ i t greater
h
attention to the process of marshaling more effectively
aU those mathematical tools built into our educational programs, we
can increase the value of mathematics in our professional practice.
The example presented in Section VI1 is used to make the more
abstract discussion concrete. Take time to compare its form to the
steps listed in Table 1. Notice how the disciplined use of these steps
facilitates the process of discovery.

I wish to express my appreciation to Henry D. Bowen, who stimulated my interest in this subject and helped me appreciate the magic
inherent in modeling.

ASAE. 1967. Yearbook.

Bowen, H. D. 1966. Developing tractable problems from abstract problem
situations. ASAE Paper 66-510. St. Joseph, MI: ASAE.
Carslaw, H. C., and J. C. Jaeger. 1959. Conduction of Heat in Solids, 2nd ed.
London: Oxford Univ. Press.
Fry,
1941. Industrial mathematics. Am. Math. M o n t h l ~ 1-38, Part 11
of the June-July Supplement.
Johnson, W. C. 1944. Mathem~ticaland Physical Principles of En~neeringAnalysis, New York McGraw-Hill, pp. 206-217.
Nahikian, H. M. 1964.A Modern Algebra for Biologists. Chicago, IL: Univ. Chicago Press, p. 2.
Noble, B. 1967.Applications of ~ n d e r ~ a d u Mathematics
a~e
in ~ n ~ n e e r i nNew
g.
YorkTheMathematicalAssociation
of America and TheMacmillan
Company.
Novak, J. D., and D. B. Gowin. 1984.Learning How to Learn. New York Gambridge University Press.
Prat, H. 1952. Can. 1, Bot.
39.
Ver Planck, D. W., and B. R. Teare. 1954.~ngineeringAnalysis: An ~ n t r o ~ ~ c ~ i o
to Professional eth hod. New York Wiley pp. 229-250.

Modeling has attracted the attention of many system analysts because

it helps to predict the behavior of real-world systems without disturbing actual systems. Many models developed are either spatial or
temporal owing to objective constraints or some limitations faced by
the analysts (Chen, 1983; Chen and Hashimoto, 1980; Doorenbos and
Pruitt, 1977; Driessen, 1986; Fluck et al.,1992a, 199213; Hashimoto,
1984; Hill, 1983; Hitch, 1977, Hoffman, 1979; Hutber, 1973; Jensen et
al., 1971, 1990; Jones et al., 1991; La1 et al., 1991, 1993; Negahban et
al.,1993; Panesar etal., 1989; Papajorgjietal.,1993;Parikh,1985;
Parikh and Kromer, 1985; Searl, 1973; Smith, 1992). But the behavior
of the systems predicted by these spatial/temporal models has both
spatial and temporal characters. Thepredicted behavior of the model
does not duplicate the actual behavior, and one of the major reasons
listed for nonduplication of the actual system behavior is the lack of
3

Panesar

either spatial or temporal character. Thus, these models have limited

application.
Analysts and researchers have always wanted to enhance the
capability of these spatial/temporal models so that the actual behavior of the system could be predicted with greater reality. The main
aim of this chapter is to familiarize students with methodology that
can be used to add the missing character in the spatial/temporal
models. Thechapter discusses the spatial and temporal characteristics
of a system, the certainty of model output, time-space functions and
their utility in developing methodology for adding the missing character, and integration methodology and gives an example of an energy model where the integration methodology has been successfully
used.

The features that vary and/or are of concern for regional planning
on the micro or macro levels are defined as spatial features. Spatial
features are taken as attributes for database management of a region.
These features help in identification and classification of homogeneous subregions in a region of interest. The spatial features are the core
of spatial models, and their accurate identification by use of a set
me tho do lo^ becomes important to eliminate any errors. Both the
variability and the equality among subregions have been used in the
literature for the identification of spatial features (Bennett and Tan,
1981; Nelson, 1987; Negahban et al., 1993;Panesar et al., 1989; Pari&,
1985; Pari& and Kromer, 1985; Smith, 1992). Thetwo approaches are
dual, i.e., equality is the dual of variability and vice versa.
It is a known fact that spatial variability exists for almost all
levels of systems, for example, in a ~ i c u l t u ~production,
al
transport,
industrial production, and the environment at themacro level and in
nutrient transport in soils and nutrient uptake by plants and inbiogas
digesters at the micro level. Spatial variability is of concern to planners and modelers because they want equal performance from or
equal distribution of benefits to spatially variable subregions. This
leads to the concept of equality among subregions. Hence, spatial
variability and equality concepts can be used to delineate homogeneous subregions or areas however small these subregions might be.
Spatial variability will help in identifying some characteristics of the
region, and equality might help to identify some other spatially variable characteristics of the region. For example, in allocating inter-

and
l n t e ~ r a t i nSpatial
~

~emporal

governmental grant resources, the equality concept will help in identification of the tax base as a characteristic fora region that otherwise
might be skipped.

. T

All natural and man-made systems are dynamic. The dynamism in

the system adds the time dimension as the behavior of a system is
different at different times. System features that are due to the time
dimension are referred to here as temporal features. The behavior of
rice production or a processing system differs with time and hence
has temporal features. The two most widely discussed and talked
about temporal features are stability and dynamism. These are discussed next.

A system has stability if it exhibits stable behavior from one period

of time to another. If rice production in the United States is the same
in 1991, 1992, 1993, and 1994, then the rice production system is said
to have exhibited stable behavior. It is an important feature from the
point of view of planners and forecasters that they do not have to
worry about the complexity of a stable system. The stability feature
of the system is good for short-term planning, but this should be used
very carefully for long-term planning as it may lead to faulty decisions and actions.

Temporal models are dynamic and hence require that past and future
behavior of the system be considered in any analysis. The past is of
utmost significanceforbiologicalsystems,for
these systems have
memory. The future has great importance for man-made nonbiological systems such as industries, transport, roads, electricity supply,
sewerage, and water supply. Man-madesystems using biological systems as components require that analysts take into account both the
future and past behavior of the system. The future behavior would
determine how the system would meet the future requirements of the
society for which it was made. All these factors add dynamism to
both natural and man-made systems, which complicates the task of
system analysts and modelers.

The certainty in model output means that once the model output is
predicted for an event it (the output) does not change when the
event actually occurs. Thus model outputs shouldhave a minimum
of uncertainty. Uncertainty, the dual of certainty, is due mainly to
four factors: (1) the stochastic nature of model inputs, (2) the stochastic determination of model parameters that are otherwise deterministic, (3) the deterioration of system behavior over time, and
(4) uncertainty about model structure. The stochastic nature
model inputs is self-explanatory. The model parameters are deterministic, but the methods (experiments, instruments, etc.) used to
determine them make them stochastic, For example, the spring constant of a spring is a deterministic feature of the spring, but the
technique used to determine it makes it stochastic. The spring constant is ete ermined by measuring force and displacements with the
help of instruments (which have errors); plotting force vs, displacement, and finally determining slope with the use of regression analysis makes the spring constant a stochastic parameter.
Deterioration in a systems performance, i.e., changes in the
systems behavior, with time is not uncommon. For example, the
field capacity of a soil at a given location is a constant parameter.
Rut the soil structure may change due to such activities as puddling
and inundation of water, which are part of the rice production system. Thus, the field capacity of soil changes with multiple use of the
rice production system, which therefore results in uncertainty in
model output.
The model structure of a system is based on knowledge available at the time the system was modeled. Advances in scientific
knowledge are being made at a rapid
pace. These scientificadvances
might change the model structure, thereby creating uncertainty in
model outputs at later
a date. The certainty feature of temporal models is of utmost importance to planners and forecasters.

.
Spatial and temporal features can be derived from time-space
functions. The following discussion indicates how time-space functions can leadtospatial/temporal
models and how missing
temporal/ spatial characteristics can be added to spatial! temporal
models.

A function of the following form is called a time-space function:

where
= output set for a time-space model
= space set
y, z for three dimensions and

y for two-

dimensions)
= set of model parameters
t = time
Let t = T where T is a specific time. Then, from Eq.
(l),the spatial

is a spatial model wit

as its outputs and P' as the model parameters. The parameter
ncludes the effect of time period T. Thus,
predicting system behavior from a spatial model over different time
haracter to a spatial model.
pace.Then,fromEq.
(I), the temporal
(3)

as its output and

ludes the effect of
n different homog
a region with the help of a temporal model will add spatial character
'l, it is inferred that some or all model
parameters of a spatial model would change with T whereas some
ode1 parameters of a temporal model would
accumulates the effect of a time period while
the effect of spatial variability in a region.
Equation (3) illustrates how temporal models lose spatial variability and hence are not fit for regional equality. Equation (2) illustrates how spatial models lose dynamism and stability, and hence are
not good for predicting the future behavior of the system.
In the literature, mathematical forms of spatial and temporal
models have three basic characters: (1)model output, (2) model inputs, and (3) model parameters. Model coefficient terms have also
been used in place of model parameters in econometric models. However, we use model parameters in this chapter. Inmathematical form,
the model output is written either as a function of model inputs,

anesar

parameters, and outputs or in the form of a di~erentialequation involving time and/or space derivatives of the model outputs along
with the model inputs, parameters, and outputs. Either the model
input or model parameters or both are functions of time or space or
of both time and space. Thus, in temporal models, some or all of the
model inputs are functions of time, and in some cases timemay also
appear as a model input, but model parameters are not a function of
time. However, both Parameters and inputs of temporal models
might be functions of space. In spatial models also, similar to temporal models, some or all model inputs are functions of space variables, but model parameters are not functions of space. However,
both parameters and inputs of spatial models might be functions of
time.
If we compare Eqs. (2) and (3) in light of the above discussion,
it can be inferred that P and P consist of two distinct subsets of
model inputs and parameters. The members of the subset of P or
that are functions of space and time, respective
model inputs; and the members of the subset of
functions of space and time, respectively are referred to as model
parameters in more general terms.
~
~ The ~modified
~ expression
~
for
Z crop water
e
requirements
without water stress (CWm)to include spatial character is (Panesar,
1993):

K ET0 + I
where ET, is the reference crop eva~otranspiration, m/cropseason;
K is the crop factor; and I is the total in~ltrationloss, m/crop season.
CW,, is the model output.
The first factor in the estimation of CW;, is the crop factor K,
which depends on the crop, season of planting or sowing, growth
stage, and weather conditions such as wind speed and relative humidity. Doorenbosand Pruitt (1977) gave values of K for several crops
during four stages of crop development, i.e., initial stage (approximately 10% ground cover), crop development stage (ground cover
increases from 10% to about 70--80%), midseason stage (crop cover
remains nearly complete), and late season stage (crop cover decreases
from h11 to harvest or maturity crop cover). They outlined a procedure to determine crop factor K for each day after planting. Smith
(1992) considered water demand during land preparation and nursery raising but only for rice. Waterdemand during the two
operations
is important for other crops also. Thus, factorK has spatial and temporal variations. Hence, K will form part of the model inputs.
ET,

oraland Spatial
Integrating

99

The second factorin the estimation of CW,, is the reference crop

evapotranspiration (ET,). Several methods, such as pan evaporation,
temperature, radiation, and combination methods (Jensenet al., 1990),
have been used to estimate ET,. The Penman-Monteith approach, a
combination method, is most widely used as this approach represents
a basic general description of the process of evaporation from vegetation and considers surface roughness and canopy rather than any
specific crop. Smith (1992) gave the following equation based on the
Penman-Monteith approach and asrecommended in the FAO Exper
Consultation held in May 1990 in Rome.
d
- (ET,)
(ET,)
dt

900
T + 273
lOOO[A + y(l + 0.34U)I
+

where ET, is the reference crop evapotranspiration, m/crop season;

R, is the net radiation at crop surface, MJ/(m2*day);G is the heat flux
from the ground, MJ/ (m2*day); is the latent heat of vaporization of
water, MJ/kg; T is the mean daily air temperature, "C; U is the average daily wind speed measured at 2 m above the ground; e, is the
saturation vapor pressure of water vapor at air temperature, Wa; ed
is the saturation vapor pressure at the dew point of the air, Wa; A is
the slope of vapor pressure curve for water, Wa/K; y is the psychrometric constant, Wa/K; t is time in days; and 900 is a constant in
kg.K/kJ. The factors R,, G, T, U, A, and ed have spatial and temporal variations. However, the factors y, and 900 are not functions
of time and space. Hence, the factors R,, G, T, U, e,, ed, A, and f will
y, and 900 will be model
be model inputsand the factors
parameters.
The third factor in estimation of CW, is the infiltration loss I for
wet cultivated crops. Soil is thoroughly worked and its structure destroyed during wetland preparation for planting. This operation is
commonly known as puddling. Two main aims of the puddling operation are to loosen the soil to facilitate transplanting and to reduce
the infiltration rate. The infiltration loss was estimated on a daily
basis using the equation
I = I,&

where I is the loss of water due to infiltration, m/crop season; I, is

the loss of water due to infiltration from saturated undisturbed soil,
for the infiltram/crop season; and ki is the soil reduction factor (>l)
tion rate. Both I, and ki have spatial variations and therefore are categorized as model inputs.

anesar

The model for crop water requirements without water stress has
output
Inputs
Parameters

cwxn
K ; R,, G, Tt U,
900

ed,

4 t, Io, ki

The inputs are functions of time and space as discussed earlier. Therefore output (CWm)is also a function of time and space in addition to
the parameters listed above. Thus, the model output ( C ~has
~ a)
form similar to Eq. (l). The exact mathematical functional form for
inputs with time and space is not known. However, only an implicit
functional relationship in theform of time-space tables for each
input
exists, Thereforethe model has to be implemented in the form given
by Eqs.
Panesar (1993) has developed the model as a temporal
model by defining CW,, ETo, I, and Io in m/ day and removing the
time derivative from Eq. (5). Hence, the Panesar model does not provide spatial crop water requirements with water stress.

Most models found in the literature are classified as either spatial or

temporal. These models have limited application as they do not reflect the system behavior in both time and space. Researchers have
mostly developed either spatial or temporal models due to certain
limitations and hence have always felt the need to expand the scope
of both types of models to predict both the spatial and temporal behavior of the system. The method of predicting the spatial and temporal behavior of a system from either a spatial or temporal model
is called here an i n t e ~ a t i o nme tho do lo^. Figures 1 and 2 illustrate
the i n t e ~ a t i ome
~ tho do lo^ for temporal and spatial models, respectively, to predict the combined spatial and temporal behavior of
a system. The integration methodology requires the design of input
and output modules. The input module segregates the spatial and
temporal data so that inputs and coefficients of the model with both
spatial and temporal variations are made available to the model. The
model will give outputs that will be stored and inte
output module to predict both the spatial and temporal behavior of
the modeled system. The model is run a number of times for different
time periods to add temporal characteristics to spatial models and
for different homogeneous regions to add spatial characte~~tics
to
temporal models. Complete knowledge about the system, the model,
and assumptions; the type of data formats for model inputs, parameters, and outputs; and language for computer implementation of the

lntegrat~n~
Spatia~and Temporal Models

MODEL INPUTS
PARAMETERS
REGION-1

REGION- 2

SPATIAL
AND
TEMPORAL
DATA

MODULE OUTPUT
MODEL OUTPUTS

FOR
REGION. 1

REGION 2
REGION - 3
REGION 4
~

SPATIAL
AND
TEMPORAL
OUTPUT

FIGURE1 Integrating spatial aspects into temporal models.

model are prerequisites for the design and implementation of input

and output modules.
The integration methodology is a multistep process.The sequential steps are
Revisit the system that has been modeled.
Study the modeling methodology used.
Understand the assumptions and the model structure.
Note model inputs, outputs, and parameters.
Study the units and dimensions of model inputs, outputs,
and parameters.
6 . Find values of the model parameters used for computer
implementation of the model.
7. Study the language used for the implementation of the
model on the computer.
8. Find computer data formats for inputs, outputs, and parameters of the model used for its implementation.

1.
2.
3.
4.
5.

anesar

FIGURE2 Integrating temporal aspects into spatial models.

9. Design input and outputmodules, and implement them on

the computer in conjunction with the model.
10. Present model results in such a way as to facilitate
interpretation.

Only the design and implementation of input and output modules

need further discussion, as the other steps are self-explanatory. The
design of the input/ output module is model-specific, but the technique, as described below, is quite standard.

Design of the input module consists of spatial and temporal data

manipulation for use by the spatial/temporal models. Two common
types or formats of model outputs are maps and tables. The design
of the input module must also consider these desired types or formats
of the model output for better interpretation of the modeled results.
I m p l e ~ e n t a ~ oofn the input module on a computer comprises a

lntegrati~gSpatial and Temporal Models

structured program for the task and testing and debugging of the
structured program modules are needed to eliminate programming
errors. The spatial input module adapts spatial and temporal data for
the spatialltemporal model such that the model processes all data.
The adapted data of the system will be in the form of several files or
subdirectories, each representing either a specific period or a specific
homogeneous region. The fileor subdirectory with model inputs and
parameters is processed by the spatial/temporal model. Hence, the
input module carries out two tasks: (1)data adaptation of inputs and
parameters into files or subdirectories and (2) processing of adapted
data from the file or subdirectory.

Design of the output module consists of manipulation of the output

from spatial/temporal models so that the missing character is added
to the system behavior and on to the model output. As discussed
earlier, two common types or formats of the model outputs are maps
and tables, The design of the output module must therefore consider
the desired type or format of the model output to provide for better
interpretation of modeled results. The implementation of the output
module comprises structured programming for the task and testing
and debugging of the structured program modules to eliminate any
programming errors. The output module should provide the capability of storing the model output for a specific period or homogeneous region in a file or subdirectory. The processingof these output
files orsubdirectories is done in such a way that the missing character
is added to the predicted behavior of the system by the spatial/temporal models. Hence, the output module has two tasks: (1)to store
the model output of a specific period or homogeneous region in a file
or subdirectory, and (2) to integrate the output files or subdirectories
in order to predict both the spatial and temporal behavior of the
system in a desired fashion.

A spatial energy model fot agricultural production (SENMAP) was

developed by Panesar (1993). The model is depicted in Fig. 3. The
model predicts requirements in terms of water, diesel fuel, electricity,
labor, and total energy for crop cultivation over a period of 1 year,
The model is driven by a large number of huge databases for Punjab
(a state of the Union of India). These databases are on crops, imple-

anesar

I
1

ments, soils, energy coefficients for other operations, mechanization,

and irrigation. Panesar used the concepts of inte~ationfor both temporal and spatial models. SENMAP consists of three spatial component models: Water and Energy Requirements, Energy Requirements
for Soil Working Tools, and Energy Requirements for Other Operations. The spatial model for water and energy requirements is driven
by another temporal model on crop water requirements without water stress (CW,). The output of the CW, model was used by the
spatial water and energy requirements model. But the CW, model is
a temporal model that requires weather data and rate of infiltration
I in addition to crop coefficients K as discussed earlier, and these have
both spatial and temporal features. The inputin terms of
CW,
needed for the water and energy requirements model must have a
spatial character, but the output from the CW, model (which is the
input to that model) does not have a spatial character; it has a temporal character as discussed earlier.The CW, model was implemented in Fortran. Panesar (1993) designed input and output modules as illustrated in Fig. 1 to adapt inputs, parameters, and outputs
of the temporal model. The spatial and temporal data on weather, K
and I, were in several ASCII files such that each file had temporal
data of a homogeneous region. Panesar designed an input module
and implemented it in Fortran to process data for differentcrops during their cropping period and for different homogeneous regions. He
designed the output module to store model output in ASCII files so
that each file represented a subregion and to process these files to
create a CW, database for the water and energy requirements model.
He implemented the output module in Fortran and dBASE so that a
spatial database for water requirements without water stress is created for different crops to be used by the spatial water and energy
requirements model.
The SENMAP model gives spatial output for a period of 1 year
in the form of maps and tables (Panesar, 1993). Thesample format of
the model output in the form of a table is given in Table 1, where
each row is for a subregion and each column is for a model output.
Similar tablesformatted for average, maximum,and minimum values
of the coefficients for diesel, electricity, water, labor,and total energy
requirements were also presented by the model. Maps showing
regions of high, medium, and low requirements of diesel, electricity,
water, labor, and total energy also formed part of the presentation of
modeled results for quickerappraisal. A method similar to Fig. 2 was
developed by Panesar to add temporal character to the model output.
He designed an inputmodule for adapting spatial and temporal data

anesar

Table 1 Modeled Total Energy and Resource R e ~ u ~ e m e nfor

ts
Punjab State in 1991
Diesel

Electricity
(GWh)

Labor
of
man-hr

Water
(ha m)

Energy
(GJ)

46,809.68
34,738.44
49,619.08
58,852.33
31,102.42
17,866.85
r 35,997.36
20,632.91
48,323.08
41,174.13
12,458.31
49,350.39
446,925.10

186,508.60
27,325.54
42,922.50
114,302.70
109,266.20
96,769.38
181,792.30
107,014.30
248,312.10
190,691.10
40,538.72
202,080.50
1,547,524.00

396,871.2
248,950.8
409,184.4
467,543.5
317,750.7
184,199.6
417,121.5
140,911.0
439,159.9
372,796.1
152,361.5
210,476.7
3,757,328.0

638,287
733,073
471,335
714,324
266,147
215,237
361,110
235,944
513,610
384,488
106,591
575,977
5,216,123

11,373,274
7,435,317
11,019,501
13,000,204
6,368,922
4,442,695
9,485,957
4,652,270
13,220,959
12,339,237
2,545,513
14,459,249
110,343,072

District
Amitsar
Bathinda
Faridkot
Ferozpore
Gurdaspore
Hoshiarpore
~a~a~da
Kapurthala
Ludhiana
Patiala
Rupnagar
Sanpr
State totals

Source: From Panesar,

such as weather, area and yield of different crops, cultural practices,

irrigated area, area irrigated by tubewells, and mechanization (tractors, draft animals, diesel-operated tubewells,
electricityoperated tubewells, etc.) for use by SENMAP. He implemented it in
Fortran and dBASE such that the temporal character could be added
to the outputs from SENMAP. The need for two l a n ~ a g e s / e n ~ i r o n ments arose from the fact that a complex calculationtask can be easily
and time-efficiently handled by Fortran and database manipulations
can be easily and efficiently done in dBASE. Thus the key in the use
of computer language for implementation of the input module is ease
and time-efficiency. Use of proper and suitable language makes the
~plementationeasy, efficient, and productive.
Panesar (1993) did not design a complex output module for
computer ma~pulationof spatial and temporal results. The output
module carried out only one task-it stored the model output. It did
not process spatial and temporal output. Panesar used computer manipulation of the spatial/temporal outputs from the model outside
the model, and hence the ~anipulationwas restricted to representing
model output in a tabular form to display spatial and temporal characteristics. Table 2 shows the modeled results with temporal character

Models
Temporal
~ n t e and
g r a t Spatial
i~~

107

Relative Total Energy and Resource Requirements for

Punjab State During the Period 1980-1991
(1980 = 1.000)
Diesel
Electricity
Energy
Water
Labor
District (GWh)
(kL)
Year
Amritsar
Bathinda
Faridkot
Ferozpore
Gurdaspore
Hoshiarpore
Jallandar
Kapurthala
Ludhiana
Patiala
Rupnagar
Sangrur
State totals

1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1992
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991
1980
1985
1991

Source: From Panesar,

h)(man.
1.ooo
1.331
1.445
1.000
1.259
1.339
1.ooo
1.266
1.350
l .ooo
1.312
1.472
1.ooo
1.530
1.521
1.ooo
0.848
1.028
1.ooo
1.380
1.540
1.ooo
1.220
1.408
1.ooo
1.283
1.398
1.ooo
1.329
1.415
1.ooo
1.289
1.347
1.ooo
1.406
1.641
1.ooo
1.300
l .424

1.000
1.370
1.617
1.ooo
1.659
1.849
1.ooo
1.678
1.693
1.ooo
1.355
1.531

1.278
1.392
1.000
1.068
1.547
1.ooo
1.551
1.911
1.237
1.461
1.ooo
1.756
2.093
1.ooo
1.345
1.517
1.ooo
1.333
l .544
1.ooo
1.953
2.556
1.000
1.460
1.740

1.ooo
1.299
1.408
1.182
1.195
1.ooo
1.287
1.374
1.ooo
1.387
1.617
1.ooo
1.462
l.424
1.ooo
0.850
1.021
1.ooo
l .200
1
1.ooo
1.l45
1.302
1.ooo
1.086
l .068
l .ooo
1.224
1.223
1.ooo
1.l42
1.111
1.ooo
0.990
0.882
1.ooo
1.199
1.263

(ha-m)

(GJ)

1.ooo
1.370
1.617
1.ooo
1.660
1.850
1.ooo
1.677
1.692
1.ooo
1.355
1.531
1.000
1.279
1.392
1.ooo
1.068
1.547
1.ooo
1.551
1.912
1.ooo
1.237
1.461
1.ooo
l .755
2.092
1.ooo
1.345
1.518
1.000
1.334
1.546
1.ooo
1.955
2.559

1.000
1.876
2.070
1.ooo
1.696
2.002
1.ooo
1.608
1.824
1.ooo
1.649
1.922
1.ooo
1.951
1.915
1.ooo
1.035
1.579
1.ooo
l .722
1.922
1.ooo
1.241
1.463
1.ooo
1.476
1.583
1.000
1.622
1.762
l .ooo
1.677
1.606
1.ooo
1.689
2.070
l .ooo
1.619
1.833

1.491
1.733

anesar

added to the model output. The results show relative diesel, electricity, water, labor,and total energy requirements in relation to 1980, yet
the table effectively displays the spatial and temporal behavior of the
agricultural production system.
The spatial output from the S E N ~ A Pmodel is in the form of
either maps or tables, Thus the outputmodule could also be designed
and implemented in Fortran, dBASE, and ARC-INFO such that the
spatial changes over differenttime periods could be displayed in the
form of maps (results of one period overlaid over the results of another period to pinpoint differences) or in the form of tables with
columns covering different periods and rows representing different
regions. TheFortran language would be very useful for complex calculations, such as finding totals, computing results for presentation
in tabular form, and getting hard copy of result tables. The dBASE
language/ environment helps in transferring model output to attributes of various homogeneous regions. ARC-INFO helps in developing maps for the model output, overlaying maps of different time
periods for visual display of changes, and getting hard copy of these
maps. Hence, implementation of the output module in a combination
of several small modules formulated in a suitable computer language
(which may differ among modules) may be helpful and easy for better presentation of modeled results.

Researchers and system analysts have developed either spatial or

temporal models for the system that actually have both spatial and
temporal characters such as spatial variability or equality and temporal stability or dynamism. ~ncertainty in
model output limits the
application of the model, and uncertainty of model structure is one
of the reasons for the behavior. Lack of either spatial or temporal
character in temporal or spatial models adds uncertainty to model
structure. Thus, researchershave always felt the need to integrate the
missing character into both spatial and temporal models.
Spatial and temporal models are derived from the time-space
functional relationships within the system. Temporal models are derived from time-space functions for a specific region,whereas spatial
models are derived from the same time-space ~ n c ~ i o for
n s a specific
time period. The concept of time-space functions helps in building
a methodology for integrating the missing character into spatial and
temporal models.

The integrating methodology is a sequential step method comprising revisiting the system that has been modeled; studying the
modeling methodology used; understanding the assumptions and
model structure; noting model inputs, outputs, and parameters;
studying units and dimensions of model inputs, outputs, and parameters; finding values of the model parameters used for computer implementation of the model; studying the language used for the implementation of the model on the computer; finding computer data
formats for inputs, outputs, and parameters of the model used for its
implementation; designing input and outputmodules; implementing
input and output modules on the computer in conjunction with the
model; and presenting model results to facilitate interpretation. The
input module is designed and implemented for two tasks: data adaptation of inputs and parameters into files or subdirectories for use
by the model and processing of adapted data from the file or subdirectory. Similarly the output module is designed and implemented
for two tasks: storing the model output of a specific period or homogeneous region into a file or subdirectory and integrating the output files or subdirectories to predict spatial as well as temporal behavior of the system in a desired fashion.
The integration methodology has been demonstrated by taking
an example from an agricultural production system that has both
spatial and temporal character.

Bennett, R. J., and K. C. Tan. 1981. Space-time models of financial resource

allocation. In:Dynamics Spatial Models. D. A. Criffith and R. D. M a c h n o w
(Eds.). Proceedings of the NATO Advanced Study Institute on Dynamic
Spatial Models, Rockville, MD: Sijthoff and Noordhoff.
Chen, Y. R. 1983. Kinetic analysis of anaerobic digestion of pig manure and
its design implication. A g i c . Wastes 855-81.
Chen, Y. R., and A. G. Hashimoto. 1980. Substrate utilization kinetic model
for biological treatment processes. Bio~ec~noZ.
Bioeng. 22:2081-2095.
Doorenbos, J., and W. 0. Pruitt. 1977. Crop 'Water ~ e ~ u i r e m e n FAO
t s . Irrigation
and Drainage Paper No. 24. Rome: FAO of the United Nations.
Driessen, P. M. 1986. Nutrient demand and fertilizer requirements, In: Modelling of A ~ i c u l t u r a lProduction- W e a t ~ eSoils
~ and Crops. H. van Keulen
and Wolf (Eds.). Wage~ngen:PUDOC, pp. 182-200.
Fluck, R. C. 19132a.Energy analysis of agricultural systems. In:Energy in Farm
Production. Energy in World Agic~Zture,Vol. 6. R. C. Fluck (Ed.). Amsterdam, The Netherlands: Elsevier Science, pp. 45-52.

11

Panesar

Fluck, R. C. 1992b. Energy conservation in agricultural transport. In: Energy

in Farm Prod~ction.Energy in World Agriculture, Vol. 6. R. C. Fluck (Ed.).
Amsterdam, The Netherlands: Elsevier Science, pp. 31-36.
Fluck, R. C., C. Fonyo, and E. Flaig. 1992a. Land-use-based phosphorus balances for Lake Okeechobeesubbasins. Appl. Eng, Agric. 8(6):813-820.
Fluck, R. C.,B. S. Panesar, and C. D. Baird. 1992b. Florida Agricultural Energy Consumption Model (FAECM).'Unpublished report. Gainesville, FL:
Agricultural Engineering Department, Institute of Food and Agricultural
Sciences, University of Florida.
Hashimoto, A. G. 1984. Methane from swine manure: Effect of temperature
and influent substrate concentration on kinetic parameter ( K ) .Agric. Wastes
9~299-308.
Hill, T. H. 1983. Simplified Monod kinetics of methane formation of animal
wastes. Agric. ~ a s t e 5:l-16.
s
Hitch, C.J. 1977.Modeling E n e r g y - E c f f n f~nteractions:
f~~
Five Ap~roaches,Washington, DC: Resources for the Future.
Hoffman, L. 1979. Energy demand in developing countries: Approaches to
estimation and projection. In:Procee~ingsof the W f f r ~ h oon
p Energy Data of
Develo~ing~ f f ~ n t r i eVol.
s . I. Paris: International Energy Agency, Organization for Economic Cooperation and Development, pp. 109-118.
Hutber, F.W.1973. Modelling energy supply and demand. In: Energy Modelling. Special Energy Policy Publication. Guildford, UK IPC Science and
Technology Press, pp. 4-32.
Jensen, M. E., J. L. Wright, and B. J. Pruitt. 1971. Estimating soil moisture
depletion from climate, crop and soil data. Trans. A S A E 24:954-959.
Jensen, M. E., R. D. Burman, and R. G. Allen (Eds.). 1990. ~vapotranspiration
and ~ ~ i g a t i oWater
n
~ e ~ ~ i r e m e nNew
t s . York American Society of Civil
Engineers.
Jones, J. W., E. Dayan, L. H. Allen, H. V. Keulen, and H. Challa. 1991. A
dynamic tomato growth and yield model (TOMGRO). Trans. A S A E 34(2):
663-672.
Lal, H., C. Fonyo, B. Negahban, J. W. Jones, W. G. Boggess, G. A. Kiker, and
K. L, Campbell. 1991, Lake Okeechobeeagricultural decision support system (LOADSS). ASAE Paper No. 91-2623. St. Joseph, MI: American Society
of Agricultural Engineers.
Lal, H., G. Hoogenboom, J. P. Calixte, J. W. Jones, and F. H. Beinroth. 1993.
Using crop simulation models and CIS for regional productivity analysis.
Trans, A S A E 36(1):175-184.
Negahban, B,, C.Fonyo, W. Boggess, J. Jones, K. Campbell, G.Kiker,
E. Hamouda, E. Flaig, and H. Lal.1993.LOADS: a GIs-based decision
support system for regional environmental planning. Proceedings of the
Conference on Application of advance^ ~nfor~ation Technfflff~es: E~ective
Manage~entof ~ a t u r a 2Resources, St. Joseph, MI: American Society of Agricultural Engineers.
Nelson, R. 1987. State-space modelling of residential, commercial and peak
demands. J. Forecast 6:97-115.

Spatial Integrating

111

Panesar, B. S. 1993. Spatial Energy Model of A ~ i c u l t u ~ Pro~uction.

al
A Ph.D.
Dissertation. Ann Arbor, MI: U.M.I.Dissertation Services, A Belt and Howell Company.
Panesar, B. S., A. P Bhatnagar, J. K. Parikh, and J. P. Painully. 1989. Modelling
of Energy andA ~ i c u l f ~~nteractions:
re
A Case Study of P u n j a ~State. Ludhiana,
Punjab, India: School of Energy Studies for Agriculture, Punjab Agricultural University.
Papajorgji, l?.,J. W. Jones, J. P. Calixte, and C. Hoogenboom. 1993. A generic
geographic decision support system for policy making in agriculture. In:
Proceedings of the 13-14 Dec. 1993 Conf. on Integrated Resource Management and Landscape Modifications for Environmental Protection. St.
Joseph, MI: ASAE, pp. 340-348.
Parikh, J. 1985. Modeling energy and agriculture interactions. I: A rural energy interaction model. Energy 15(7):793-804.
Parikh, J., and R. Kroner. 1985. Modelling energy and agriculture interactions. 11: Fuel-fodder-food-fertilizer relationship for biomass in Bangladesh. Energy 15(7):805-817.
Searl, M. F. 1973. Energy Mo~eZing.Washington, DC: Resourcesfor the Future.
Smith, M. 1992. CXOPWAT-A C o ~ p u t e rpro^^^ for ~ r r i g a t i oPlanning
~
and
Manage~ent.FAO Irrig. & Drain. Paper 46. Rome: Food and Agriculture
Organization of the United Nations.

This Page Intentionally Left Blank

c.

Linear programming is a modeling tool that can assistin thesolution

of many problems in agriculture. In particular, linear p r o ~ a m m i n g
is useful in selecting the best alternative from a number of available
courses of action.
For example, consider the situation where a farmer is planning
for the upcoming growing season. She has three crops she is considering planting on her 200 hectares (ha) of land. Considering all factors, the farmer estimates that 300 h of labor time can be devoted to
these three crops. Assuming that machines and capital are not limiting, the farmer wants to know how many hectares she should plant
in each crop in order to maximize her profits. She receives $50 profit
per hectare for crop $35 for crop 2, and $40 for crop 3.

Sowell and ~ a r d

11

This example illustrates the type of problem linear programming

can assist in solving. The farmer has to decide how much land to
allocate to each crop. She wants to maximize profits but has limited
land and labor. There are many combinations of cropping plans she
can choose from, but she wants to know which will yield the most
profit.
To begin to model this problem into a linear p r o ~ a m m i n gformat, several d e ~ ~ t i o are
n s needed. A linear programming problem
consists of three major components:
l. Decision variables
2. Objective function

3. Constraints
The decision variables relate to the decision that must be made and
are expressed with algebraic symbols. In the above example, the decision variables (xj,j = i, 2, 3) can be defined as

xi = number of hectares of land that should be planted in

crop j if profits are to be maximized
The objective function is an expression of the measure of effectiveness against which alternatives will be compared. The alternative
with the optimum (maximum profit or minimum cost, for example)
is the best alternative. In the crop planning example above, the
objective function will be established by first defining the profits for
each crop j , on a per-hectare basis, as cj. The objective function can
then be expressed as
Maximize z = clxl

+ c2x2+ c,x3

(1)

or
3

j=

The value of the objective function is expressed in dollars because

the units on cj are dollars per hectare for crop j and the units on the
decision variables (xj)are hectares of crop j . Thus the combination of
that yields the largest z is the optimum cropping plan.
.The objective function could be increased without limit if there
were no constraints on the problem. In this case, the farmer has only
200 ha of land and 300 h of labor time to devote to the crops in
question. The constraints must be expressed in algebraic equations.
In the case of the constraints, the amount of each resource needed to

odel ling with Linear ~ r o g r a m m i n ~

15

produce one unit of output must be known. The term is used to

describe the amount of resource i needed to produce one unit of
product j . The term bi is used to represent the total amount of resource
i available. Thus, the constraints for a linear programming problem,
with j = 1,2,3 and i = 1,2 (as in theabove example), canbe expressed
as

The less-than-or-equal-to signs in these equations stem from the fact

that the land andlabor limits are not to be exceeded. In other words,
the farmer may use less than the maximum of one resource if the
optimum solution deems it best to do so.
If, in the example, it is assumed that each hectare of crop
requires 5 labor hours; each hectare of crop 2 requires 3 labor hours,
and each hectareof crop 3 requires 4 labor hours, then the total linear
programming formulation of the crop planning example canbe stated
as follows:

Subject to:
x1

+ xz + x3 S 200

+ 3xz + 4x,

300

(land constraint)
(labor constraint)

and

The latter constraint (xi2 0, i = 1, 2, 3) is associated with the simplex

method for solving linear programming problems, but in this case it
also has a physical interpretation in that there will be land either
planted or not planted.
Identifymg and formulating linear p r o ~ a m m i problems
n~
is as
much as an art as it is a science. The more one formulates linear
programming problems, the better ones ability to formulate will become,The following examples serve to introduce students to the
range of problems in agriculture that can be solved with linear programming and further their understanding of the terms and terminology used in formulating linear programming problems, especially
the concepts of decision variables, objective function,
and constraints.

EXA

The crop mix problem described in Section I is often more commonly

referred to as a product mix problem. In the situation described,
the farmer had a number of crops being considered for production
and a limited amount of resources. The question was, How much of
each crop should be grown in order to maximize profits? This example is, of course, rather simple (e.g., small numbers of crops and
resources considered) in order to illustrate the concept. The number
of crops may always be limited due to equipment availability,but the
number of different resources can often be quite large in a realistic
problem.

The diet, or feed mix, problem deals with determining the amounts
of alternative feeds to provide for livestock. The objective is to provide feeds that meet certain nutritional requirements at minimum
cost. The decisionvariables are often the amounts of the various feeds
to be included in the diet. The objective function coefficients define
the cost per unit of each feed so that in minimizing the objective
~ n ~ t i one
o n is minim~ingthe cost of feeding the livestock. The constraints express the need to meet minimum ~utritionalrequirements.
Each nutrient has its own constraint.
Consider the situation where a farmer, among his several enterprises, raises hogs for market. He has three types of feed available
for his hog operation and is mainly concerned about nutrients. He
has had each feed analyzed to determine the amounts of the three
nutrients in a pound of each feed. He has also determined the
imum daily requirements of each nutrient for each hog and the cost
per pound of each feed. He now wants to know how much of each
feed should be fed to each hog if he is to meet the mini mu^ nutritional requirements at a minimum cost. The data he has obtained are
given in Table 1. Each entry under feed has units of nutritional
value per pound of feed.
The decision variables for this problem can be defined as

xj = pounds of feed j to be fed daily to each hog

W i l e the cost of each pound of feed is known (from Table l),the
objective function can be written as
Minimize z = 12x,

+ 6x2 + 8x3

odeling with Linear ~rogramming

17

Table 1 Data for the Diet Problem

Nutrient

Daily nutritional
requirement

Feed
Corn Ib"

Tankage lb-l

Alfalfa Ib-l
___

1
2

150
50

cost

60
55

5
12

18
40
8
6

40
75
12

Each nutrient will have a constraint that expresses the fact that
its daily requirements must be met. Thus, the constraints are

60x, + 18x2+ 40x3 2 150 (nutrient 1 constraint)

55x, + 40x2 + 7'5x3 2 260 (nutrient 2 constraint)
5x, + $xz + 12x, 2 50 (nutrient 3 constraint)
and, as before, x,, xz,and x3 must be nonnegative.
This example illustrates how inequalities (2)and minimum objective functions can arise in linear programming problems.

A soybean producer grows soybeans in four different fields located

around the county. She stores her beans at two different sites. With
rising energy costs, she wishes to determine the amount of beans
from each field that should be stored at each storage area if transportation costs are to be minimized.
The first storage location has a 7'25 m3storage capacity, the second storage facility can hold 510 m3, and the trans~ortationcosts are
estimated to be $0.095/ (m3-km). The trucks are assumed to carry a
full load on each trip.
The production of each field and the distances between the fields
and the storage sites are given in Table 2.
The decision variables for this problem can be defined as

xii = m3 soybeans shipped from field i to storage location j

The cost of shipping a cubic meter of soybeans from field to field j
is defined as

Sowell and

l1

le 2 Data for the Transportation Problem

Expected
yield
size
no.Field

Field
(ha)

m3/ha

150255
152
80
175
210

1
2
3
4

1.7
1.9
298l .7
2.1
9

Hauling distance
Total
(m3)

Storage 1

Storage 2
10
29

M1

14
6
12
18

= cost per cubic meter per kilometer [$/ (m3 h ) ] times

the distance from field i to storage location j in h

Thus,
cll = $0.095/(m3*h) 14 km = $1.33/m3

With these costs

Minimize

shipping, the objective function becomes

Constraints on the problem relate to the need to store all the

beans and thelimits on storage capacities. Thus, forthe need to store,
the following constraints apply.

xll + xI2= 255 (field 1 supply constraint)

xzl + xZ2= 152 (field 2 supply constraint)
x31+ x32 = 298 (field 3 supply constraint)
x43+ x42 = 441 (field 4 supply constraint)
Limits on storage capacities can be expressed as

xll + xzl + x31

x12+ xzz +

+
+

x43

5 725

(storage site 1 constraint)

x42

5 520

(storage site 2 constraint)

As before, the xs 0.
This formulation will allocate the beans from each of the four
fields to the two storage sites in such a way as to minimize the total
transpo~ationcost.The problem ensures that all the beans are
shipped (first constraint set) and that the amount shipped to each

storage location does not exceed its storage capacity (second constraint set).

The basic problemin linear programming is one of either maximizing

or minimizing a function (objective function) of several variables (decision variables), with the value of the variables being subject to a
number of constraints. In linear programming, both the objective
function and the constraints must be linear functions. The mathematical statement of the general linear programming problem (with
n variables and m constraints) is as follows:
Maximize
(Minimize)

z = clxl

+ c2x2 +

+ c,,x,

(3)

Subject to:
ailxl

+ ai2x2+

+ airZxn

i = 1, . . . , m

=, 2)b,

(4)

and
x1, x2,

. ,x ,22

(5)

* *

where only one of the operators (5,

=, 2)holds for each of the m
constraints. The quantities cj, aii, and bi, i = l, . . . , m and j = 1, . . . ,
n, are all known constants while the x;s are the unknown decision
variables. This general form fits the formulations developed in the
previous examples, and the student should verify this fact.
The c i s are often referred to as cost coefficients or objective
function coefficients; the a i s may be referred to as A matrix coefficients or structural coeficients; and the bj are often referred to
as right-hand-side coefficients or b coefficients.
The standard/ form of the linear programming problem will
often be referred to in later discussions. This is the specific form of
the linear programming problem statement in which all constraints
are equalities and the right-hand side of each constraint is nonnegative. Thus, the linear programming problem in standard form is as
follows:
Maximize
(Minimize) z = clxl

+ c2x2 + - + c,x,
*

Subject to:
ailxl

+ ai2x2+

+ ailzx,= b,

i =

. . .,m

and
x2,

* /

x, 2 0

The method used to solve linear p r o ~ a m m i n gproblems, the

simplex method, requires, among other things, that the linear programming problem be in standard form. This requires that constraints
expressed as inequalities be converted to equalities.

Inequality constraints are converted to equations through the use of

slack or surplus variables. A slack variable is added to the lefthand side of a less-than-or-equal-to constraint to take up the slack
implied by the S sign. On the other hand, a surplus variable is subtracted from the left-hand side of a eater-than-or-equal-to constraint
to take off the surplus implied by the 2 sign.
To illustrate, consider the constraint,
dllxl

+ a12x2 + al3x3 S

x1

To make this inequality an equation, the variable x4 will be added,

and it will assume whatever value is necessary for the equation to
hold. Thus, adding x4 to Eq.
results in
allx,

+ a12x2+

+ x4 = b,

For Eq. (10) to hold, x4, like the other variables, must be nonnegative.
From the standpointof meaning relative to theinitial problem, x4 can
be interpreted, for example, to be the resources (represented by the
constraint) not used in the problem, i.e., the slack resources.
For the constraint, consider the following:
%lXl

+ a12 x2 + 43x3 2 bl

In this case a surplus variable is subtracted from the left-hand side.

It then assumes the value needed to make the inequality an equation.
GllX,

+ a12x2 +

- x4 = bl

(12)

Again, for Eq. (12) to hold, the surplus variable, must be nonnegative. To see the physical meaning, recallthe diet example. In that
situation the surplus variable represents the nutrients supplied over
the m ~ i m required.
u ~
Slack and surplus variables, while added to convert in inequalities to equations, do have meanings with respect to the original problem and are therefore a part of the problem. They carry valuable

information about the problem that will be explored in more detail

later.
When all the inequalities that exist in a set of constraints are
converted to equalities, the set of constraints is in the form of a set
of linear equations, Thus, linear programming can be thought of as
a problem of optimizing a linear function in which the variables are
constrained to be nonnegative and satisfy a set of simultaneou
equations.

.
In trying to understand how linear programming optimizes a linear
function (objective function) in which the variables must satisfy a set
of simultaneous linear equations, it is useful to graphically view what
is happening. Since no combination of variable values that does not
satisfy the constraints can beconsidered, it is of interest to graphically
see this set of solutions and how it related to the objective function.

.
Consider the problem
Maximize z = 80x, + 90x2
Subject to:

+ x,
2x1 + X ,
3x1 + 5x2
x1

40
x,, x2 2 0

The first step in the graphical solution is to determine the set of

numbered pairs (xl,x,) that are the feasible solutions. On the xl, x2
coordinate plane (Fig. 1) the first quadrant yields the set of points
that satisfy the nonnegative restrictions x1 2 0 and x2 2 0. To find
the set of points in the first quadrant satisfymg the constraints, we
must interpret geometrically inequalities such as

x, + x2 5 10
If the equal sign holds, i.e.,
x1

+ x, = 10

we have the equation for a straight line, and anypoint on that straight

122

line satisfies the equation. Notice that any point lying in the halfplane below the line satisfies the constraint
3-

22

owever, only those below the line and in the first quadrant satisfy
both the constraint and the nonnegati~tyrestrictions. Consider the
other constraints in a similar manner, and wefind a set of points that
satisfy all the restrictions of the problem. This set of points is represented by the cross-hatched area of Fig. 1. All points in this shaded
area are feasible solutions since they satisfy all the restrictions. This
set of points is a convex set,* and it can be shown that the set of
*A set is convex if all points on a line connecting any two points in the set is also in

the
e.g.,
set,

is a convex
set.

isconvex
not a

set.

odeling with Linear ~rogramming

123

feasible solutions to a linear programming problem is always a convex set.

Now consider the objective function. To solve the linear program, we must find the point or points in the region of feasible solutions that give the largest value of the objective function. For any
fixed value of z,say zl,

is a straight line. Any point on this straight line will give a value of
= 900. For each different value of z we obtain a different straight
line, but the lines are all parallel. We wish to find the straight line
with the largest value of z that has at least one point in the region of
feasible solutions. Notice (Fig. 1)that the line

has no points in the region of feasible solutions and thus is not the
optimal feasible solution to the problem. Let z2 = 720, and note that
the line

has points in the region of feasible solutions but that z can have a
larger value and still be in the region of feasible solutions. By drawing
lines parallel to and z2 we find that the maximum value that z can
attain and remain in the region of feasible solutions is found where
the lines x1 + x2 = l 0 and 3x1 + 5x2 = 40 intersect, or at x1 = S, x2 =
5. The value of z at the point is 80(5) + 90(5) = 850. Note that the
maximum value of z occurs at an extreme point or corner of the
region of feasible solutions. It can beshown thatthe optimal solution
of a linear programming problem always occurs at an extrerhe point
of the convex set. In this example the optimal solution is unique, i.e.,
only one (xl,.x2) pair lies in the region of feasible solutionsand optimizes the objective function. The following example is one for which
there are an infinite number of optimum feasible solutions.

Lets now look at the graphical solution of the linear programming

problem
Minimize z = -4x1

+ 4x2

Sowell an

Subject to:

The region of feasible solutions (Fig. 2) is constructed in a manner

similar to that in the previous example.
Notice that the objective function

is parallel to the line

x1

x, = 4

between the lines

xz = 0

and

2x1 + 3x2 = 13

and that the minimum value of z occurs along this line. Although
this is an extreme point of the convex set, the objective function has
the same minimum value of -16 anywhere along the line

FIGURE2 Graphical solution to a multiple optima problem.

25
x1

x, =

between the lines

xz = 0

and

2x, + 3xz = 13

There are an infinite number of (x,, x,)

pairs along this line and consequently an infinite number of solutions to the problem.
Now let's turn our attention to the graphical representation of
problems for which no solutions exist.

Maximize z = Zx, - x,
Subject to:

Proceeding as before, we sketch the set of points that satisfy the

restrictions of the problem (Fig. 3). Whenwe draw a typical objective
function

we see that it lies in the region of feasible solutions. We also see that
for any other value, zzr of the objective function such that z, > z1 the
line

2x1 + x, = z,
also lies in the region of feasible solutions. This means that the objective function does not attain a maximum value on the set of points
satisfying the constraints. This set is unbounded, and the problem
has no solution. The situation is also referred to as a problem with
an unbounded solution.

126

Sowell and Ward

FIGURE3 Graphical representation of an unbound solution problem.

Minimize z = 2x1 - 3x2

Subject to:

When we drawthe straight lines for the constraints and attempt

to find the intersection of half-planes that satisfy the constraints, we
find that no region of feasible solutions exists, i.e., there are no num-

~ o ~ e ~ with
i n glinear Programming

27

*2

x1

FIGURE4 Graphical representation of a problem with no solution.

bered pairs (x,/ in the coordinate plane that satisfy all the restrictions. Therefore/ the problem has no solution. (See Fig. 4.)
Summarizing, we have looked at examples of linear programming problems for which (1)there is a unique optimal solution, ( 2 )
there are an infinite number of optimal solutions, (3) the solution is
unbounded, or (4) there is no solution.

The graphical view of linear programming portrays very clearly the

formation of the region of feasible solutions using linear equations.
It also demonstrates that the optimal solution of a linear programming problem is a solution to the set of simultaneous linear equations

derived from the constraint inequalities and equalities. Thus, Linear

programming can be viewed as mainly concerned with the solution
to simultaneous linear equations that arise from the restrictions on
the variables (the constraints). The goal is to find the solution to the
simultaneous equations for which all variables are nonnegative and
that results in the maximum (or minimum) value of the objective
function.
Through the addition of slack and surplusvariables, a constraint
set of inequalities can be converted to a set of simultaneous Linear
equations. In trying to solve this set of equations, however, there are
normally more unknowns than equations. In this case, there are a
large number of solutions.
One way to obtain a set of solutions to a system of m independent equations and n unknowns (where m n) is to arbitrarily select
n - m variables, set them equal to zero, and solve the ~emainingm
equations in m unknowns. By this method one can find solutions,
where )(: is the number of ways n things can be taken m at a time,
l.e,,
):7(

Linear p r o ~ a m m i n garrives at a solution in this manner. ln addition,

linear p r o ~ a m m ~
ensures
g
that the next solution will be at least as
in minimization) as
good (i.e./ 2 if a maximization problem and
the previous solution. Therefore in most cases the number of solutions to be considered will be less than (:).
At this stage it would perhaps be instructive to see a specific
example of a linear programming problem solved using this procedure.
I)

Suppose a farm tool manufacturer has two primary resources, machine time and labor hours. In a certain period of time he has 200
machine-hours and 300 labor-hours to devote to three products, xl,
x,, and x3. Product xl takes 15 machine-hours and l 0 labor-hours per
unit. Product x, takes 10 machine-hours and 25 labor-hours per unit,
and product x3 takes 10 machine-hours and 20 labor-hours per unit.
The manufacturer wishes to determine the mix of products that will
maximize profits while not exceeding the machine-hours available.
He wishes to provide full employment for his workforce and therefore requires that all available man-hours be used. This is a competing

objective, for the maximum profit might be greater if the manager

were not so concerned about his employees. Thetwo constraints and
nonnegativity restrictions in algebraic form are
15x1

10x2 +lox, 5 200

10x1 + 25x2 t-20~3L= 300

x1,

x,

x3

The profits to the manufacturer will be $5 per unit of xl,$10 per unit
of x,, and $12 per unit of x,. Thus, the optimizing to be performed is
Maximize 5x,

+ lox2 + 12x,

Since the first constraint is a S inequality we must add a slack variable and rewrite the set of constraints as
x1

+ 10x2 + lox, + x4 = 200

loxl + 25X2 + OX, = 300

where x4 does not need to be included in the objective function since

it has a coefficient of zero.
Since n = 4 and m = 2, there are (;) = 6 solutions that can be
obtained by setting variables equal to zero two at a time and solving
for the remaining two variables. First setting x3 and x4 equal to zero,
we get

+ lOx, = 200
10x1 + 25x2 300

15x1

=I

or x1 = 7.2727 and x, = 9.0909. In a similar manner we can obtain the

five other solutions. Table 3 shows all the possible solutions and the
value of the objective function for each.
The solutions (3 and 4) that are marked with anasterisk in Table
3 involve a negative variable, which violates the nonnegativity restriction. The objective function is not evaluated for such solutions.
These solutions are said to be infeasi~lebecause they do not satisfy
the conditions of the problem. The other solutions that do not satisfy
the conditions of the problem are said to be feasible. Solution6 is the
feasible solution that maximizes the function and is said to be the
~ a x i feasi~le
~ u ~ solutionor o ~ t i solution.
~ a ~
It must be emphasized here that by a solution to the problem
we are referring to a combination of the n variables and not to the
maximum value of the function. This is because we are basically concerned with solutions to simultaneous equations. This is logical be-

Table 3 Six Solutions to Farm Tool Manufacturer Problem

Solution
no.
1
2
4"
5
6

X1

7.2727
5
30
0
0

x2

x3

9.0909
0
0
12.5
0
0
- 20
40
12 120 0
0 180 15

x4
0
0
-250
0
80
50

Objective function
(5x1 + 10x2 + 12x3)
127.27
175

*See text.

cause the manufacturer is primarily trying to determine what products to make to maximize profits with the resources available.

Thesimplex method or algorithm is an iterative procedure that

moves from one corner of the region of feasible solutions to another,
always improving the value of z, or at least not making it worse.
Viewed another wax the simplex algorithm is an iterative procedure
that moves from one basic feasible solution (of the system of simultaneous linear equations) to another, again always improving z.At
each iteration one variable is removed from the solution and another
inserted. At each iteration the solution is tested for optimality. If optimality is reached, the iterations end. If optimality has not been obtained, a new iteration is performed. This continues until optimality
is reached or until an unbo~ndedor infeasible solution is obtained.
The manner in which variables are chosen to leave solution ensures that the next solution will be feasible (all x's 2 0), and the
manner which variables are chosen to enter solution ensures that
the next solution has the largest improvement in x possible. Optimality is reached when it is no longer possible to improve z.
The simplex method needs a basic eas sib le s o ~ ~ in~ order
i o ~ to
begin. A basic feasible solution, recall,is a solution to the linear constraint equations in which the values of the variables in solution (the
basic variables) are nonnegative. With this initial solution, the simplex
proceeds as follows.

Step I . Check the optimality of the current solution. This is

accomplished by first determining the value of zj - cj for the nonbasic

~ o ~ e l i with
n g Linear ~rogramming

31

variables. This is the relative effect of these nonbasic variables (those

not in solution) on the objective function if they were to be brought
into the solution, If none of the nonbasic variables has an improving
effect on the value of z,the optimality has been reached. The value
of cj is simply the objective function coefficient for the jth variable. zj
is defined as the product of the objective function coefficients of the
variables in solution at this iteration (cBi)and the value of the
ssociated with variable j at this iteration
where summation on i is over all variables in solution.
Step 2. If some zj - cj c 0, then the nonbasic variable having
the smallest (largest absolute) zj - cj is selected to enter solution.
Denote the column in which the incoming variable is located as k.
Step 3. The variable to leave solution. is selected by examining
the constraints to see how large the nonbasic variable can be made
before one of the current variables in solution goes to zero. The variable that goes to zero first is removed. If the variable that is to be
removed is located in the rth row, the objective of the examination of
the constraints is to find Y such that

where xBi is the value of the variables in solution at each iteration.

The rationale for this "minimum ratio rule" is discussed later. The
condition Y i k > must be observed in simplex calculations, and this
is where many students in their first simplex computations make a
mistake-be careful!
Step 4. With the variables coming in and going out decided,
the next step is to compute the new solution. This is accomplished
by "pivot operations"-a sequence
of elementary calculations that
results in a system of simultaneous linear equations being reduced to
an equation having a unit coefficient and all other variables in the
equation having coefficients of zero. Thissolution of systems of linear
equations is often referred to as the Gauss-Jordan elimination technique. The result of the pivoting operation is to have the identity
matrix associated with the variables in the solution.

To further explain the mechanics of the simplex procedure, the tableau format is now described and applied to a simple maximization

132

problem. The tableau format is basically a means of organi~ingthe

calculations and the results of the simplex procedure. Further, by expressing the pivoting operation in simple formulas, the entire simplex
procedure becomes rather mechanical, and thus readily programmable on the computer.
Consider the following linear p r o ~ a m m i problem
n~
where al,lz+l
through an,n+mare slack variables.
Maximize z = clxl

+ c2x2+

+ c,x,,

Subject to:

The ~~~~~a~
tableau, using the above notation, is shown in Table
4. In this tableau, xBirepresents the value of the variables in solution
(initially xBi = hi); zj = cBiyij,where summation on i over variables in
solution; and zj - ci represents the relative measure of effectiveness
coefficient (sometimes referred to as rj).
The initial tableau, after a few iterations, can be generalized as
shown in Table 5. In the general tableau, the are the variables in
solution at a given iteration and yij are the values of ay at an iteration
after the initial tableau.

Initial Tableau for the Simplex Method

odeling with Linear

133~rogramming

Table 5 General Tableau for the Simplex Method

CB1

cl

Yll

y12

CB2

c2

y21

yz

y1

l*

Y2n

y1,1,+1

0
y2,w+2

...
* *

0
0

The specific mechanics of performing the four previously discussed steps of the simplex procedure are as follows, where the terminology is that of the above tableaus.

tep 1. Check optimality of current solution. Check to see if all

zj

0. If they are all positive or zero, the optimality has been

reached; there is no variable that can be brought into solution and
cause z to improve.
- cj 2

Step 2. If one or more of zj - cj S 0, then the solution can be

improved by bringing in a new variable. Select the column having
the smallest (largest absolute value) zj - cj and call this column k.
The variable in the kth column (i.e, the kth variable) will enter solution at the next iteration.
Step
The variable to leave solution must be determined. As
noted, before, the variable to be removed from solution is determined
by finding r such that

The variable in the rth row is removed,

Step
The new tableau is computed by pivoting. Pivoting is
nothing more than establishing the identify matrix with the variables
in solution. As noted before, this simply means that in each column
associated with a variable in solution, there must appear one unit
coefficient and the rest zeros. Thispivoting can beexpressed through
the equation

bl
b2

13

gij = yij - yik

[~],

+ r;i =

. . ., n

+ m; i =

. ., m

where ijijand i J j are the entries in the new tableau.

The entries is the new tableau for xBiand zj - cj can be computed
by applying the above pivoting equations to those locations in the
tableau. In this case the zj - cj would be treated as ym+l,jand the
pivoting equations used. The new zj - cj could also be computed
from their definitions. In this case the zj row in the tableau is needed;
otherwise it is not needed. Likewise, the xBiare treated as yi,n+m,.land
computed using the pivoting equations.
Once the new tableau is established, the procedure returns to
step 1. Iterations of the procedure continue until either optimality is
reached, an unbounded solution is found, or an infeasible solution is
encountered. These other possible terminations of the simplex procedure are discussed later.

Consider the problem

Maximize z = 2x1 + 5x2 + 4x3
Subject to:

2x1 +x2+ x3
4x12x22x3

6
5

x11 Xzr
0
First, the inequalities must be converted to equalities by adding slack
variables. Thus, the constraints become
2x1 +

+ + =6
4x1 + 2x2 + 2x3 + x5 = 5
The initial tableau, using the above data, is shown in Table 6.
Here x4 and x5 are the variables in solution as they are readily
and x3 are set to zero. So we have an initial
evaluated when xl, xz,
basic feasible solution to begin the simplex procedure. The zj and zj
- cj were computed using the data in the tableau from the initial
statement of the problem.
Proceeding to step 1, a review of the zj - row reveals that
optimality has not been reached. All nonbasic variables (xl,x,, and
x3)have negative values, indicting that any one could be brought into
solution in step 2 and result in animprovement in the objective func-

Table 6 Initial Tableau for Simplex Example

cj

Variable
in soln

cZ3i

x1

x2

x5

x4

x5

0
-2

zj

zj - cj

xBi

0
-5

6
5

tion. The value of z2 - c2 is the smallest, indicating that bringing x2

into solution will result in the greatest improvement of z.Thus, x2 is
to enter solution at the next iteration (in the next tableau), and therefore k = 2.
In step 3, the variable to leave solution is determined by the
minimum ratio rule. The row of the variable to leave solution (r) is
determined by selecting the row with the smallest ratio of xni/yizr
where yik > For row 1,
-6=6
y12 1
and for row 2,
XBl
"-

3"
5
- = 2.5
2

y22

Thus, row 2 (r = 2) contains the variable to leave solution (x5),The

variable yZ2becomes the pivot element as this is yvkin the pivoting
equations of step 4-it is involved in all pivoting operations.
To pivot (step 4), the equations noted earlier can be used noting
that k = 2 and r = 2. Thus,each new
can be computed directly
from the equations. Forexample, g12 (with = 1 and j = 2) is
computed:
-

y12

g12= 2 - 2

(:)

Q22

y12

(E)

le 7 Second and Final Tableau for Simplex Example

CBi

Variable
in soln

x5

0
5

7/2
512

zj

5.
0

zj - cj

5/2

0
0

25/2

5/2

In a similar manner, using the appropriate pivoting equations, allnew

entries in the next tableau can be computed.
It is also instructive to note that the above equations will always
result in the identity matrix associated with the variables in solution.
Thus, the pivoting could be performed by simply performing the row
manipulations that result in the identity matrix being associated with
the variables row in solution, without memorizing the equations. The
equations do the same thing. Thus, the tableau with x4 and xz in
solution is as shown in Table 7. The zj row can be computed from its
definition, and it then is used to compute zj - cj, or the zj - cj can be
computed by simply pivoting on the values from the previous
tableau.
From the tableau of Table 7, it can be seen that all the zj - cj
entries are greater than or equal to zero; therefore, optimality has
been reached at this tableau. The optimal solution to the problem is
Maximize z = 25/2
x1

= 0;

x, =

5
-*

= 0;

x* =

7
2

-*

x5

=0

Of course, it is rare for a problem to be solved in one iteration, but

the purposes here are to illustrate the simplex mechanics.

To further illustrate the modeling and solution of linear p r o ~ a m m i n g

problems, consider the situation where a farmer has
ha of land,
400 labor hours that can be devoted to crop production on these 100
ha, and ~25,000of capital to fund the production on this land, The

farmer has the equipment to produce three different crops and wants
to h o w which crops, and how many hectares of each, to produce in
order to maximize profits. The farmer has the crop data shown in
Table 8.
The decision variables can be defined as follows.

xi= number of hectares of crop i produced

The objectivefunction represents the profit maximizationgoal if profits per hectare are multiplied by number of hectors (xi)for each crop.
Thus, the objective function is
Maximize z = 60x, + 55x2 + 68x3
subject to constraints on land, labor, and capital. Being careful to ensure that units are compatible across the inequalities, the constraints
can be modeled in terms of the amounts of each crop produced as
follows:

+ x2 +
3x1 3- 4x2 + 5x3
300x, + 260x2 + 230x3
x1

(land)

x3

400

(labor)

25,000

(capital)

Also, the nonnegativity restrictions apply, so these three inequalities

are a part of the model:
x1 22

0,

x2

0,

x3

The land constraint could be an equality if the farmer wants to use

all the land and maximize profits. However,the assumption is made
that maximizing profits is the objective subject to a limit of 100 ha of
land being available.
To solve using the simplex procedure, the constraint inequalities
are corrected to equalities via the addition of slack variables. These
slack variables represent unused land (x4),unused labor (x5),and unused capital (x6).
Data for Crop Mix Problem

Crop 1
Crop 2
Crop 3

Labor hours required

per hectare

Capital needed
per hectare

Profit per
hectare

4
5

260

60
55

The initial simplex tableau can now be established using as the initial
basic feasible solution the values of x4, x,, and x6 when x1 = x, =
x3 = 0. These variables are placed in the solution initially and the
basic solution is immediately available.

x4 = 100 ha of land
x, = 400 h of labor
x6 = 25,000 dollars of capital
Thus, the initial tableau is as given in Table 9. From this tableau, it
can be seen that optimality has not been reached. Since x3 has the
smallest zj - cj value, it is chosen to enter solution. Since x5 has the
minimum ratio, it is selected to leave solution.
The second tableau, after pivoting, is shown in Table 10. Optimality has not been reached as - cl and zz - c, are still negative.
Since x1 has the smallest zj - cj, it enters solution. In determining the
variable to leave, cannot be considered as it has a negative under
the xi column. Since row 3 has the smallest ratio, x6 leaves solution.
The third tableau, after pivoting, is shown in Table 11. At this
tableau, optimality has been obtained. The optimal solution is
Maximize z = $6222 profit

le 9 Initial Tableau for Crop Mix Problem

XBi

cBi

Variable
in soln

x3
x5

x2

x1

x4
x5

x4

100
0
25,000

x6

0
zj - cj

odeling with Linear ~rogramming

Second Tableau for Crop Mix Problem

Table
Cj

Variable
in soln

CB;

x4

0
68
0

60

55

68

x1

x2

x3

x4

x5

x6

-0.20

0
0
1

20
80
6600

0
0

5400

0.40
0.60
162

x3

x6

68z; 54.4

z; - cj

40.8

- 19.2

0.20
0.80
76

0
1
0

1
0
0

-0.6

0
0

0.20

-46
19.66
13.6

xl3i

with
ha of crop 1 grown
ha of crop 3 grown
0 ha of crop 2 grown

40.74
55.56

and with 3.70 ha of land not utilized. These hectares could not be
used because all the labor and capital were exhausted (x,and x6 have
zero values).

Up to this point in presenting the simplex method, an initial basic

feasible solution has been readily available. This has been the case
because all the problems up to now have required the addition of
Third and Final Tableau for Crop Mix Problem

Table

cBi
0
68
60

Variable
in soln
x4

x2

x3

0
0
1

0.01
0.52
0.47

0
1
0

68 zj 63.42 60
z; - cj
0

8.42

x3
x1

x4

x5

x6

1
0
0

-0.09
0.37
-0.28

-0.002
-0.004
0.006

6222
0 0.09 8.36
0 0.098.36

3.70
55.56
40.74

slack variables to all constraints. What happens if there are greaterthan-or-equal-to constraints or equations as constraints in the initial
formula~on?Forexample, suppose that in the previous problem,
land was an equality in the initial formulation. This would have
meant that x4 would not be needed. Thus, the initial basic feasible
solution would have been difficult to determine.
To easily find the first basic feasible solution needed to initiate
the simplex procedure, a ~ ~ ~ ~c ai~ ~i albare
l ~ used.
s
These are variables
that are added to the problem after all inequalities have been converted to equalities. They are therefore added to equations! They are
added simply to get a basic feasible solution and to get the simplex
under way with a minimum of effort. Since they are added to equations, however,they cannot be in the final solution at a positive level.
They have no meaning to the problem (as slack and surplusvariables
do)-they are simply added to facilitate solution of the problem.
Since artificial variables are added for solution purposes only
they must be removed from the problem. One technique used to
"force" artificial variables out of the solution to a linear programming
problem is called the Big M method.
The Big M method "forces" artificial variables out of solution
by giving them a large (Big M) positive or negative objective function
coefficient d e p e n d ~ gupon whether the problem is a minimization
or a maximization problem, respectively. In this manner, it is highly
desirable, from the standpoint of optimizing the problem, to remove
the artificial variable and not let it return to the solution.
To illustrate the procedure, consider the problem
Maximize z = 3x1 + 2x2
Subject to:

+ x2 2s 4
x1 + x2 =

2x1

Converting the constraints to equalities results in two equations but

no readily apparent basic feasible solution.

2x1

+ x, + x, = 4
x1 + x, =
x1,

x, 22

In this simple problem, it would be possible to easily determine an

initial basic feasible solution; however, in most practical problems it
is not so simple. Consequently, an artificial variable will be added to

o ~ e l i with
n ~ Linear

the second constraint to illustrate how an initial basic feasible can be

easily obtained. Thus, the constraints for the problem are now

2x1 + x2
x1 + x,

+ x3 = 4
+ x4 = l

Artificial variable x,, however, must not appear in the final solution
at a value other than zero. If it does, the solution is not feasible because the second constraint is violated!
In setting up the first tableau for solving this problem by the
simplex method, x, will be assigned a large negative objective function coefficient, --M. In this manner the simplex procedure will find
that once x, is removed from solution, it will be highly undesirable
to bring it back into solution. Thus, in the initial tableau, the objective
function coefficients will be (3, 2, - M ) for the respective variables.
The first tableau is shown in Table 12.
x3 and x4are in the initial basic feasible solution,and their values
in the solution are the right-hand-side values from the two constraint
equations, as before.The only difference is that x4 is an artificial
variable.
Since x1 has the smallest value of zj - cj, it is to enter solution
on the next iteration. The minimum ratio rule results in x4 leaving
solution. Pivoting results in the second tableau, as shown in Table 13.
Since all zj - cj are positive, optimality has been reached. With x4 out
of solution, the solution is feasible. Thus, x4 permitted the simplex
procedure to be easilyinitiated, and the large negative cost coefficient
helped ensure that, once removed, the variable would not reenter
solution. Again, if x4 had been in the optimal solution at a zero value,

Initial Tableau for Example of Big M Method

CBi

Variable
in s o h

x2

-M

zi
2;

- cj

-M
-M - 3

-M
-M - 2

0
0

0
0

-M

Sowell and Ward

142

Table 13 Second and Final Tableau for Example ofBig M Method

CBi

Variable
in soln

X1

x3

X1

Zj

0
0

zj - cj

x2

x3

x4

-2

2
1

3
3+M

the solution would be feasible, but if x, had a value in the optimal

solution, the solution would be infeasible.

There are a large number of computer routines available for solving

linear programming problems. These routines vary greatly in capability Some handle rather small problems, while others are designed
to handle very large problems. Some simply give the optimal solution, while others provide considerable additional information for
further analysis of the solution. Some routines are strictly set up to
solve only minimization or maximization problems,while others can
solve either. (By simply reversing the s i p s of the objective function
coefficients, a mini~izationalgorithm can be used to solve a maximization problem. Students are asked to verify this.)
In most routines, the user must inform the computer how many
variables there are, how many constraints and of what type, and
whether the problem is maxirnization or minimization. The objective
function coefficients, the constraint matrix coefficients, and the righthand sides are input tothe computer using the specified formats. The
coefficients may all have to be specified, or the user may be able to
enter only nonzero coefficients (an advantage for large problems with
many zero coefficients in the A matrix).
Output detail may be at the user's option. The solution only
may be printed, or additional detail may be requested along with the
solution. Many algorithms provide the option of modifymg the problem and solving again.

~ Q d e l i n with
g
Linear ~rQgramming

143

In using a computer routine to solve a linear programming problem, the student should carefully read the routines documentation.
The specifics of how the routine operates, how data are input, and
what is output are described. The understanding of linear programming obtained to this point will permit students touse most routines
to solve a linear programming problem.
URAL A P P L I C A T I ~ ~ S
~ a c ~ i n e r y a n a ~ e ~ and
e n t Crop ~ r o d ~ c t i oPro
n

This example draws on Von Bargen (1980).

A 680 acre dryland cash-grain farm called Mechanized Acresis
operated by one individual. This farm is located in eastern Nebraska
and can produce corn, soybeans, grain sorghum, and winter wheat.
A fixed complement of equipment is assumed for the crop production
and grain handling system. Supplemental hired labor is available
when needed such as at harvest time.
The objective for the crop production system analysis is to determine the acreage for each crop to maximize profit with respect to
the costs that are proportional to the acreage. Fixed equipment costs
of $26,500 per year and a fixed labor expense of $12,000 per year for
the operator are subtracted from the profit function to determine the
net profit.
The following decision variables are defined:

xl = acres of corn
x, = acres of soybeans
x3 = acres of grain sorghum
x4 = acres of wheat
x5 = dummy variable associated with fixed cost
In addition to limitations on acres of cropland available, it is
assumed that there are also limitations on time available in the spring
for row crop planting and in the fall for row crop harvesting. These
times are 90 h and 160h, respectively both at an 85% probability
level. Time available for planting and harvesting wheat is assumed
to be unlimited. There is also a minimum wheat acreage of 40 acres.
This is the minimum required to justify special wheat equipment.
In order to develop the coefficients for the constraints involving row crop planting and harvesting time it is necessary to relate
machine parameters to the operations (i.e., a six-row planter

equipped for 30 in. rows, traveling at 4.2 mi/h, with a field efficiency
of 0.6 is assumed). The coefficients must have units of hours per
acre, or the inverse of field capacity. In the case of the six-row corn
planter, the coefficient is
8.25
-speed width field efficiency 4.2
= 0.22 h/acre

mi/h

15 ft

0.6

All other coefficients in the planting and harvesting constraints are

determined in a similar manner.
Based on the above information and the original statement of
the problem, the following constraints can be written:

x1 + xz + x3 + x, 5 680
0.22~1+ 0.19~2+ 0.20~3 90
0.60~1+ 0.57Xz

x1,

xzr

(land constraint)
(row crop planting time
constraint)

+ 0 . 5 5 ~=S ~l60

(row crop harvesting time

constraint)
x4 40
(wheat acreage constraint)
x5 = 38,500 (fixedcost constraint)

x31

x,, x5

22

(nonnegativity restrictions
constraint)

The coefficients for all decision variables except x5, fixed cost,
represent the net profit per acre. For each unit of x5 the net profit will
be reduced by one unit; consequently its cost coefficient is -1. The
objective function for this problem is
Maximize z = 8 4 . 5 0 ~+~9 3 . 9 0 ~+~9 3 . 0 0 ~+~4 5 . 1 0 ~-~x5
Solving this problem we find the maximum net profit to be
$6102.54. The optimal crop mix is 291 acres of grain sorghum and
389 acres of wheat.

This example is from Hilliez and Lieberman (1974).

A certain farming organization operates three farms of comparable productivity. The output of each farm is lixnited both by the
usable acreage and by the amount of water available for irrigation.
The organization is considering three crops for planting that differ

primarily in their expected profit per acre and in their consumption

of water. Furthermore, the total acreage that can be devoted to each
of the crops is limited by the amount of appropriate harvesting equipment available. The data for the upcoming season are given in Table
14.

In order to maintain a uniform workload among the farms, it is

the policy to use some of the usable acreage as conservation acreage
or wildlife habitat, and the percentage of usable acres saved or set
aside should be the same on each farm. This percentage is to be
determined by the other constraints and this one, However, any combination of the crops may be grown at any of the farms. The organization wishes to know how much of each crop should be planted
at the respective farms in order to maximize expected profit.
To begin the modeling of this problem, it is quite clear that the
decision variables, xii (i = l, 2, 3; j = 1, 2,
should be the number of
acres of the ith crop grown on the jth farm. Therefore, the objective
is to maximize total profit in dollars, where

subject to x+ 2 0 (for i = 1, 2, and j = 1, 2, 3) (no negative acres)

and the restrictions modeled below.
The restrictions on usable acreage at each farm are

+ x31

+ x21

x3l
x12

x13

+ x32

+ x22

+ x23

400

I
600
x33

I
300

The restrictions on water availability are

The restrictions on crop acreage are

x11

+ x12 + x13

I
700

x21 + x22 + x23

~ 3 % x32

I
800

+ x33 S

To model the policy of a uniform set-aside acreage, the total percentage of planted acreage per farm must be equal for each pair of the
three farms. Therefore, the equations

Data for Crop Irrigation Problem

Farm

Water available
(acre ft)

Usable acreage

1500
2

900
-

Maximum
Water
(acre
acreage

Crop

consumption
ft / acre)

Expected
profit
($/ acre)

5
4
3

x12 + x22 +
600
x11

x32

x13

400
300
100

+ x 2 3 + x33
300

+ x21 + x 3 2 - x13 + x23 + x33

300

400

must be satisfied. Since the first two equations imply the third, the
third equation can be omitted from the model. Furthermore, these
equations are not yet in a convenient form for a linear programming
model, as all of the variables are not on the left-hand side. The final
forms of the uniform workload restrictions are

+ x21 + x31) - 2(xn + x22 + x32) = 0

(x12 + x22 + x 3 2 ) - 2(x12 + x23 + x 3 3 ) =

3(x1,

The optimal solution of this problem is as follows:

xI1 = 300 acres;
x12= 200 acres;
xZ2= 250 acres;
x23

= 225 acres

Maximize z = $342,500

This example is taken from Loftis and Ward (1980).

Many government programs for conservation in agriculture involve payments to farmers for a part of the cost of conservation struc-

~ Q d e l i n with
g
linear ~rQgramming

ture or practices. Two necessary policy decisions in such programs

have to do with how a given budget is to be divided between alternative conservation measures and what fraction of the total cost of
each measure should be funded by the government.
If one is willing to make the following qualifymg assumptions,
this problem may be modeled as a set of linear constraints on resources and a linear objective function and be solved by linear
programming.
1. The objective of the program is to maximize regional benefits within given levelsof government and/or total expenditure.
2. Each dollar spent on a given conservation measure returns
a h o w n and constant dollar amount in regional benefits.
3. The program will be sufficiently attractive to farmers that
all of the government monies included in the budgetwill be
spent.
Theobjective function is therefore a linear function of the total
amounts spent on all conservation measures included in the program.
The constraint set is largely determined by political considerations
and may include any or all of the following.
l. A budgetary limitation on the total size of the program or
on the government share
2. Maximum and minimum values of the cost-share fractions
for each conservation measure
3. Limitations on the amount spent on certain measures in relation to others

An example of the last type of constraint would be a requirement

that twice as much be spent on nonstructural measures as on structural measures.
Consider a hypothetical situation in which a federally funded
program is being developed to reduce erosion and sediment yield in
the Great Plainsregions through four conservation management practices. The practices under consideration are listed in Table 15 along
with maximum and minimum cost-sharing fractions and expected
benefits of each.
The total size of the program is to be limited to $400 million,
and the federal share is not to exceed $100 million. The program is
also limited by the following political constraints:

le 15 Description of Alternative Practices in Cost-Sharing

Conservation Problem
Federal share
Conservation practice
1. Planting and managing conservation crops
0.5
2. Managing
existing
rangeland
Construction of sediment
reservoirs
4. Construction of small onfarm measures such as
terraces, ponds, and
grassed watenvays

Max
Min

'Expected benefit
($ benefit/$ spent)

2.00

0.2
0.3

0.7

1.75

0.1

0.4

2.25

0.2

0.8

2.50

At least twice as much must be spent by the government

on practices 1 and 2, nonstructural measures, as on practices 3 and 4, structural measures.
Federal expenditures for either one of the nonstructural
measures must not exceed three times the federal expenditure for the other nonstructural measure.
Federal expenditures for either one of the structural measures must not exceed three times the federal expenditure
for the other structural measure.
These constraints ensure that nonstructural practices are favored
over s t ~ c t u r apractices
l
and that there is a fairly equitable distribution of funds within each of the two categories. Limitationson spending for ~ d i v i d u a practices
l
could be imposed as well but are not
included in the present example.
Eight decision variables are required for this problem, Let
= federal

amount spent on conservation crops

x2 = amount spent by farmers on conservation crops
= federal amount spent on range manage~ent
= amount spent by farmers on range management
x5 = federal amount spent on sediment reservoirs
= amount spent by farmers on sediment reservoirs
= federal amount spent on small on-farm structures

eling with Linear

xs = amount spent by farmers on small on-farm structures

Convenient units for the x's are 1 unit = $10 million. The objective
function coefficients are defined as follows:
cj = the regional return indollars per dollar spent on decisionj;

...,

j=1,2,

The objective function is then

= Clx1

c2x2c3x3

+ c4x4 + c5x5 + c6x6 + c7x7 + csxs

The numerical values for the cost coefficients as found in Table 15

may be put into the above expression to yield

+
225x6 + 2.5OX7

z = 2.0~1 2.0~2 1.75~3+ 1.75X4

+ 2.25X5

2.5ox8

The linear programming algorithm is to maximize the above objective

subject to the constraints described earlier, which may be formulated
as follows.
The total program must not exceed $400 million:
x1

+ + x3 + x5x4

+ xx76

x2

+ x8

The federal share is not to exceed $100 million:

x1

+ + x5 + x7 I10

Political constraint (a) above:

x1 + x3

2x,

+ 2x7

or

x1 - x3 + 2x5 + 2x7

Political constraints as described in (b) above:

and
"3x1

x3

Political constraints as described in (c) above:

x5 - 3x7

and
"3x5

+ x7

Sowell and

15

An additional set of constraints ensures that the cost-sharing

fractions will remain within the allowable limits. The federal costshare fraction for the alternative (Table 15) planting and managing
conservation crops may be expressed as x,/ (x1 + x2). Since the minimum value of this fraction is 0.2 (from Table 15), the following constraint is imposed.
X1

x1

+ x2

2 0.2

which may be rewritten as

0 . 8 ~ 1 02x2 2 0

The maximum value of the same cost-share fraction is given as 0.5.

The corresponding constraint is
0.5~1 O.5X2

Similar constraints that exist for the other three conservation measures are written below.
For managing existing rangelands,
0 . 7 ~- ~0 . 3 ~
2 ~0
0.3~~

-0.7
0~~

For construction of sediment reservoirs,

0.9~5- 0.1~62 0
0.6~
- ~0 . 4 ~S ~0

For construction of small on-farm conservation measures,

0.8x7 - 0 . 2 ~ ~0
0.2~
-~

After adding nonnegativity conditions,

0
the LP formulation is complete. Thefollowing is the optimal solution:
x1,x2,

X3, X41

= 5.00;

x5 = 2.19;

x51x6x,7fx8

x2 = 5.00;
x6 = 19.71;

x3 = 1.67;
x7 = 1.14;

x4

I=

x8

0.72
4-57

m mum objective z = 87.74.

From these results the optimal distribution of funds and costshare fractions were found to be as shown in Table 16. At the optimum, both of the budgetary constraints were tight, and a total benefit
of $877.4 million would be expected from the program.

odeling with ~ i n ~ a r

l51

Table 16 Optimal Distribution of Funds and Cost-Share Fractions

Practice
Conservation cropping
2. Range management
3. Sediment reservoirs
On-farm structures
Total

Total expenditure,
optimal
($/million)

share,Federal
optimal

23.9
219.0

50%
70%
10%

57.1

20%

From a practical viewpoint, one might ask, If the above problems were developed from an actual government conservation program, would the administrator allocate the money as noted in the
solution? The answer to this question depends upon how well the
LP formulation, with all its assumptions, describes the real-world
problem. The administrator, without any further work on the LP solution, must try to evaluate (somewhat subjectively) the validity of
the assumptions, the probability that the linear programmin~coefficients will remain the same over time, and other factors not included
in the model (e.g., the likelihood that all structural measures will be
used in appropriate situations and adequately maintained and the
likelihood that nonstructural measures will be adequately maintained). Such an evaluation might cause the administrator to modify
the linear p r o ~ a m m i n results
g
somewhat in deciding the appropriate
funding levels between alternatives and the fraction of the cost of
each measure to be funded by the government.
It is possible, via sensitivity analysis, to examine the effects of
changes in the assumptions or benefit estimates on the solution. If
the model results are greatly affected by small changes in benefit
estimates or by removal or modifications of assumptions, then the
administrator would need to have confidence in the validity of his
assumptions in order to have confidence in the results.

.
This example is taken from Sowell et al. (1980).
A field with initial elevations as shown in Table 17 is to be landformed for improved water ~anagement.Based on soil characteristics and crops to be grown, it has been determined that in its final

52

Data for Land Forming Design Problem

Original field elevations"
Cross-row direction
Row direction
9.97
9.75
"Elevations are determined at stations on
the row and cross-row directions.

ft grids, i.e., spaced

9.76
ft apart in both

design the row slope of the field should be in the range of 0.1-0.5%
and the cross-row slope in the range of 0-1% for optimum water
management. The shrinkage characteristics of the soil are such that
the volume of cuts in the field should exceed the volume of fills by
a factor of 1.25-1; a tolerance of k0.05 is permitted. It is desired to
form the field to meet the above requirements while minimizing the
amount of earth moved and consequently the cost of land forming.
There are several types of land forming designs depending upon
whether constant or varying slopes are desired and where cross-row
drainage is desired (Sowell et al., 1980). For this example we choose
the design that assumes constant slopes in both the row and crossrow directions.
Formulation of the landforming design problem as a linear program is based on the assumption that the volume of cut or fill around
a field station is a linear function of the depth of cut or fill at that
station. The objective of a land form in^ design is to form the field to
meet the specified range of slopes with a minimum of earth movement, which results in minimum cost. Correspondingly, the linear
p r o ~ a m m i n gobjective function is expressed in terms of minimizing
the sumof cuts in the field. Beforegoing further with the fo~mulation
of the linear program it is necessary to define the notation that will
be used in the model.
xij = cut at station j ] ,
i = 2, j =
yii = fill at station j ] ,
i = 2, j =
U = slope in the row direction, ft/ 100 ft
v = slope in the cross-row direction, ft/

2, 3
2, 3

ft

53

Using this notation, the objective function is

It was previously stated that fills and cuts are directly proportional;
therefore, fills are not included in the objective function for they will
be minimized when cuts are minimized.
There are five groups of constraints in the linear programming
formulation of the land forming design problem. The first two are
similar, one for the cross-row direction and one for the row direction,
and represent the majority of the constraints. They are based on the
arithmetic of the differences between the elevations of two adjacent
field stations after the land has been formed. (Recall that the stations
are spaced 100 ft apart in both the row and cross-row directions.)
Consider the first two stations in the first cross row of the example
(Table 17). The difference between elevations will be equal to the
slope between the two stations, i.e.,
(10.01 - x11

+ yzl)- (10.12 - x12 + Y12) = v

The firstpart of the equation in the parentheses represents the design

i.e., the original elevation minus the cut at
elevation at station [l, l],
the station plus the fill at the station. Obviously, either the cut or the
fill will be equal to zero. Thequantity in the second set of parentheses
represents the design elevation of station [2, l].There is a similar
constraint representing the difference in elevation of all adjacent stations in the cross-row direction and all in the row direction. Rearranging terms so that the decision variables are on the left-hand side
of the equation and constants on the right-hand side, the first two
groups of constraints are as shown in Fig. 5.
In the first two groups of constraints, the row slope U and crossrow slope v are decision variables. Itwas indicated earlier that these
variables have upper andlower limits. Thethird groupof constraints
impose those limits.
Group 3 constraints are the limits on row and cross-row slopes:

0.1;

0.5;

1.0

It is not necessary to have the constraint U 2 0, since this is one of

the nonnegativity restrictions.
The fourth group of constraints ensures that the ratio of cuts to
fill is
1.25 -+ 0.05

"

"

$ 4

"

P J ?

ct'?

"

$ $

"
I

7
"

+
"

. c $

l?

a,

S
bi

P
i?

'2

~ o ~ e ~ Xi it^
n gLinear ~rogramming

155

Table
Design Cuts and Fills for Land Forming Design
Example"
Cross-row direction

Row
direction

-0.06
3-0.12

--0.17

3-0.18

0.00

2
-0.08

0.00

"Positive number indicates fill, negative indicates cut.

These constraints would be as follows:

and
i=l, 3
j=l,3

i=l, 3
j=l,3

i=l, 3
j=l,3

j=1,3
j=l,3

The fifth group of constraints are the nonnegativity restrictions.

The solution to the land forming design problem, involving original field elevations as shown in Table 17, is given in Tables 18 and
19. The design slopes in the row and cross-row directions are 0.105
ftfl00 and 0.109 ft/lOO ft, respectively. The cuts or fills at the individual field stations are given in Table18, and the final design
elevations are given in Table

Table
____~

Design Elevations for Land Forming Design Example

Cross-row direction
Row
1
2,

10.19
10.08
9.98

10.08
9.97
9.87

9.97
9.86
9.76

Sowell and War

Hillier, F. S., and G. J. Lieberman. 1974. e er at ions Research: Pri~ci~Zes

and
Practice. New York Wiley.
Loftis, J. C., and R. C. Ward. 1980. Optimal cost sharing through linear programming. ASAE Paper No. 80539. Presented at the 1980 Annual Meeting
of ASAE, San Antonio, TX, June 15-18.
Sowell, R. S., R. C. Ward, and L. H. Chen. 1980. Linear programming: Basic
concepts, simultaneous equations and graphical solutions. ASAE Paper
No. 80-5032. Presented at the 1980 Annual Meeting of ASAE, San Antonio,
TX, June 15-18.
Von Bargen, K. L. 1980. Linear programming applied to a machinery management and crop production situation. ASAE Paper No.80-5037. Presented at the 1980 Annual Meeting of ASAE, San Antonio, TX, June 1518.

Agricultural Research Service,

Stoneville, Mississippi

Department of Agriculture,

Much research in agricultural process management has focused on

the development of simulation models. These models are used primarily to test the effect of changing inputs on one or more output
state variables. Depending on the process involved, simulations vary
in their absolute accuracy and rarely achieve it under most conditions. Input parameters are difficult to properly characterize, and
complex processes may
require significant simplification.For all these
reasons simulations are used primarily in the research environment
to test different management schemes in a relative sense.
Recent research has focused on a new method to permit simulations to be used as management tools. An approach documented
herein uses readings frarn sensors to provide adjustments to input
parameters of a simulation. The model was essentially parameterized
"on the fly." The hope was that the model could make better estimates of temporal water status as readings from sensors provided
157

T~omso~

appropriate feedback. Of course, sensors cannot accomplish these adjustments by themselves. A bit of intelligence derived from field experience and keen practical knowledge of the processes involved
must be programmed into the system.
The following questions could be raised. Why might it be necessary to combine sensor- and model-based approaches? Why not
dispense with themodel and use sensors with appropriate algorithms
as the primary i n s t ~ m e n t sof decision making? The latter approach
is appropriate in many cases. For example, high volume instruments
( W I ) for cotton grading use sophisticated sensing systems that are
continually being improved upon (Thomasson, 1993). Although the
output could be used in a larger simulation of machinery sequences,
nothing needs to be modeled to determine grade because sensing
systems by themselves can do the job and do it well.
Soil water sensors, used for years to help the farm manager
schedule irrigations, are inexact in their characterization of crop water
stress. In their simplest mode, these sensors placed in the soil can
help the farmer determine when to irrigate on the basis of a preset
"trigger point," which usually indicates soil water potential or the
amount of energy the plant exerts to extract water from the soil. Soil
water sensors have been around for years, but farmers are still reluctant to use them for a variety of reasons. Some types of sensors respond well to temporal changes in water content but require servicing. Other types do not require servicing but respond more slowlyto
changes in actual soil water content. For proper irrigation scheduling,
a sensor must "track" the soil water conditions in situations requiring
frequent replenishment of water as in the dripirrigation of tomatoes
(Smajstrla, 1985). Low frequency irrigation systems such as those of
the center pivot type can, however,be scheduled with sensor readings
that cannot respond immediately to soil drying (Thomson and
Thread@, 1987). This is because sensors can be read once a day (in
the morning) for successful control.
Realizing limitations of soil water sensors,researchers have
spent much time refining crop water use models for actual irrigation
scheduling. These models need to be calibrated for soil type and crop
characteristics in a specific en~ronment.Sensed weather data are input to the models in many cases. Accuracy of critical input parameters becomes more critical as the required interval to read sensors
becomes shorter,as with crops requiring frequent replenishment. Another problem is that many of these models are site-specific or need
to be recalibrated in new environments. Periodic soil sampling or soil

rocess S i ~ ~ ~ a t i o ~

15

water sensing is wise to ensure that the model stays on track, but this
can be labor-intensive.
In situations like the above where there may be imperfections
in both sensor- and model-based management methods, a new
method of combining the two has shown promise. Sensor responses
may give clues as to which model input parameters require adjustment-a sort of dynamic calibration. Algorithms can be in place to
regulate the amount of adjustment, interpret responses to the adjustment, and take appropriate action. Ideally sensors would not be
needed as the system "learns" and provides more accurate control of
the process.
The purpose of this chapter is to summarize the development
of the method and the results from a study that used sensors and a
water use model together. Other potential applications for the method
are discussed as well as potential problems in developing the method
for a particular application.

Irrigation management can be accomplished using properly interpreted readings from sensors placed in the soil, water balance models,
or both. As has been mentioned, both methods have drawbacks. Realizing this, an objective of combining the two to provide better control was implemented.

The model chosen for the study was PNUTGRO 1.02 (Boote et al.,
1989a). PNUTGRO is a process-oriented peanut crop growth model
that is an adaptation of a soybean model,SOYGRO (Wilkerson etal.,
1983; Jones et al., 1988). The model predicts dry matter growth, leaf
area index (LAI), crop development, and final yield depending on
daily weather data and specific soils. Primarily developed as a research tool, PNUTGRO allows scientists to test the effect of different
soil, weather, cultivar, and management factors on crop yield. The
model can be runon a time step of 1 day covering the entire growing
season. Water balance components of this model were implemented
in the system described herein.
The knowledge system described here could be used with any
crop simulation model that uses the IBSNAT (1986) format for input
and output files. In addition to PNUTGRO, models that follow the
IBSNAT format include CERES-MAIZE (Jones and Kiniry 1986),

CER~S-WHEAT(Ritchie, 1985),and SOYCRO (Wilkerson et al., 1983;

Jones et al., 1988). PNUTGRO 1.02
was chosen because lysimeter evaluations were performed in this study on Florunner peanuts.

Two weighing lysimeters (300 220 cm) were inst~mentedto determine the soil water balance. Both lysimeters were installed at the
Irrigation Research and Education Park (IREP) by Butts (1988). For
this study each lysimeter was instrumented with four load cells, a
net radiometer, and thermocouples. Peanuts (Flomnner)were planted
by hand on May 20, 1988 in a moist soil cm (1.5 in.) deep and
cm (1.5 in,) apart in each row. The plants were thinned to cm (3
in.)apart in thenorth lysimeter when a suitable stand was observed
15 days after planting (PAPL), June 3. Plants in the south lysimeter
lagged somewhat behind the ones in the north lysimeter in the early
stages and were thus thinned 121PAPL, June 9.
The e~uivalentof560 kg/ha of a 5-10-15 mix of fertilizer was
incorporated into the soil for each lysimeter. A decision was made
not to incorporate a preplant herbicide because the lysimeters were
small enough to be weeded manually, Watermark* soildater sensors
(Irrometer, 1989)were placed at several depths (Fig. l) on June21 in
one location of each lysimeter that showed a good stand population.
These sensors were read manually at least once a day using the Watermark 30KTC meter. To obtain soil water tension from meter readings, a curve was derived for the Watermark 30KTC meter (Thomson,
1990). This curve was related to the sensors ac resistance, whichis a
function of soil water tension and temperature (Thomson and Armstrong, 1987). To provide temperature input or compensation of the
soil water sensors, thermocouples were placed
20, 40, 60, and 90
cm below the surface, the same depths as the sensors. Since some
sensors were placed under the growing canopy and some under exposed soil early in the season, two banks of thermocouples were
placed in each lysimeter to represent both covered and exposed conditions. Thermocouples and load cells were read every half hour.
A manually activated sump pump and dual tank system were
used to drain the lysimeters after large rain events. The quantity of
water removed was measured and subtracted from the total water
removal indicated by the load cell measurements. In this way water
Use of trade names in this publication does not imply endorsementby the authors
of the products namedor criticism of ones not mentioned.

IGURE

1 Soil water sensor placements in the root zone.

use by the crop during the period of sump pump operation could be
estimated. Details of calibration procedures for the load cells and a
description of the data acquisition system are given by Thomson et
al. (1989) and Thomson (1990).

The knowledge-based irrigation management system consists of five

parts. These parts and the programming language or shell used to
develop each part are
1. An expert system for sensor evaluation (VP-Expert)
2. Data input and calculation routines (Pascal)
3. Parameter adjustment routines applied to PNUTGRO input
files (Pascal)
4. PNUTGRO codemodifications (Fortran)
5, Irrigation scheduling routines (Pascal)
Some parts of the system are dependent on others and were thus
developed concurrently. Each component is detailed below.

An expert system to interpret soil water sensor readings was developed from previous knowledge about soil water sensor responses to

l6

Thomson

wetting and dqnng as described in detail by Thomson (1990). The

principal objective was to flag possible errors in sensor readings so
they would not be used to adjust model input parameters. Decisions
were based on
1. Whether or not enough d y n g had occurred to compare sensor responses and thus allow evaluation
2. Readings that may be out of the sensors calibrated range of
operation
3. Twe of sensor (ranges of operation may be different)
4. A sensors physical proximity to other sensors
5. A sensors response to drying compared to that of other sensors at the same depth

Figure 2 briefly illustrates the expert systems use in interpreting

sensor readings. These data were taken from Watermark granular matrix sensor readings performed in a lysimeter study. Figure 2a shows
responses at the 8 cm depth both in the row and 15 cm away from
the row. (Readings away from the row were used to gain knowledge
about lateral root distribution.) Sensor readings at the same depth in
the center of the row of each lysimeter show a slight temporal divergence. The expert system, however, allows a certain amount of variability and both sensor readings in the row would be accepted in
this case. Thus,readings in the row were averaged and could be used
for parameter adjustment.
The bottom graph, Fig. 2b, shows a drying response at a much
deeper location. Mostof the readings shown are below the lower limit
of tension (l0 Wa) at which the sensors can be read accurately
(Thomson and Armstrong, 1987; Pogue,1991). If the sensors were
tensiometers, readings would be accepted because they do not show
an out-of-range condition at the low end of their operating range.
Out-of-range conditions at the high end of operation are not presently
treated in the knowledge base, although rules could be developed.
(Tensiometers can break tension at high tensions.) Thus, it is assumed that the soil is not permitted to dry beyond the workable
range of the sensor before irrigation. Although an out-of-range reading cannot be used to adjust soil properties, the program still uses an
out-of-range reading that is increasing to infer water use in the deeprooted horizons. Along with readings from other sensors, relativesensor responses can infer a root distribution and be used to adjust root
weighting factors.
Daily sensor data were input to a Pascal routine (FIELDRAW)
(Fig. 3) and imported by a spreadsheet macro into the spreadsheet

rocess Simu~ation

Julian Day

163

1
"""""""..
l

Sensor

Parameter Input

FIGURE3 Data flow for parameter adjustment method.

S E N R E ~ . W DATRUL-l/
~.
written inthe rule development language of VP-Expert, evaluated sensor data taken from the spreadsheet and placed flags
directly
into
the
spreadsheet file,
~ E ~ R E ~ next
. W to~ readings
/
that were questionable. Readings
that were out of the sensor's range of operation received a 1 in the
flag location. Sensors that were judged to be giving false readings
based on criteria 4 and 5 above were given a 2. Different flags were
placed forthese two situations because SENPOS, the parameter modification programf still used out-of-range readings that were increasing to modify a root weighting function WR(I), which is discussed
later. Absolutevalues of sensor readings that were out of the sensor's
range of operation could not be used for other adjustments. Spread-

sheet macros in SENRED were written to import ASCII sensor data

from FIELDRAW, move data within the spreadsheet so DATRUL-l
could analyze a new set of sensor readings, and print data out to an
ASCII file for use by SENPOS in adjusting PNUTGRO parameters.

To construct the knowledge base for PNUTGRO parameter modification, routines were developed to modify the PNUTGRO weather
file, accept soil water tension data from the user, extract soil water
contents and root length densities from PNUTGRO output files, convert soil water contents to tensions, and display the results.
The user is asked to enter sensor data or run the system with
default sensor files. If the default option is chosen, the sensor data
acquisition routines are bypassed. If the user has sensor readings, the
program asks how many sensor stations are in the field and requests
values of the previous days total solar radiation, maximum and minimum air temperatures, rainfall, and photosynthe~callyactive radiation (PAR).These are necessary inputs to the PNUTGRO crop
model. If the user does not have a reading for PAR, the program
calculates an approximate value from solar radiation. (PAR is needed
in photosynthesis routines and influences specific leaf area. PAR is
not used directly for water balance calculations but does influence
root growth and root water uptake.)
If the program is being run for the first time, the user is asked
for the number of sensors being read at each station and for sensor
depths and lateral spacings in sequence. The program presents the
user with a graphical representation of the input sensor layout for
approval. Thisallowscorrection of mistakes the user might have
made in entering the layout. The user is then asked for the Julian
day time of day the sensors were read, and tension values for each
sensor in sequence. The program stores each stations readings in a
separate data file. Each data file keeps the latest 2 days of sensor
readings.
M e n all sensor data have been entered, the program sends
weather data to the input weather file for PNUTGRO at Julian day
1. The program uses weather data for a typical high stress day for
several days inthe future to forecast the earliest date irrigation should
be required (Ross, 1984). Irrigation trigger level is based on a userentered soil water tension value that is converted from water content
using a programmed soil water characteristic curve.
The number of future days that use weather data for irrigation
forecasting is variable. After the last forecast day data for a typical

T~omso~

season are entered into the datafile, If data without rain were placed
in the PNUTGRO weather file for the entire season, the crop would
show si@ficant stress during each run of PNUTGRO. Although the
crop stressing later in the season would not affect the short-term irrigation forecast, other useful results (such as yields) could not be
obtained. The program is thus designed to allow other PNUTGRO
variables such as yield to be examined as desired using a prior
weather file and daily updating.
To make later comparisons between soil water contents based
on sensor data andon the crop model, the program weights soil water
content and root length per unit area in the defined PNUTGRO soil
horizons to zones around thefield sensors. The program must do this
because PNUTGRO soil horizons will not be at the same depths as
zones that sensors represent. Even if the user consistently placed sensors at the midpoint of defined PNUTGRO horizons, those horizons
would not be wide enough to accommodate the field sensor's zone
of influence. Each fieldsensor has an assumed zone of influence that
is automatically expanded or contracted depending on itsproximity
to other sensors in the zone. The initial zone of influence was set to
14 cm based on experience and can be changed by the user.
A procedure in the program also performs error checking if the
user enters questionable weather data. If total radiation or maximin
temperatures are outside a reasonable range, the user is alerted to
take corrective action. This may involve checking the sensor or instrumentation. Under these conditions, the weather file is not modified until the problem is fixed. If the value for PAR is questionable,
the user is alerted, but the weather file is modified with a value calculated from the total radiation. If the user enters zero rainfall for
that day but observed an appreciable amount of rain, he or she is
told to check the data collection system or strume en tat ion. In this
case, the weather file is still mod3ied if rainfall was the only questionable input value.
eter A ~ j u s t m e ~
Routi~es
t

The parameter adjustment algorithms were developed for soil characteristic and root distribution functions based on my knowledge and
the experience of two other specialists. PNUTGRO was modified to
accept ET data from the field test so differences in sensor- and modelbased results could be attributed to different soil-water and rooting
characteristics only. Complete program listings and further details on
all PWTGRO Fortran code modifications conducted for this study
can be found in Thomson (1990).

rocess Sim~lation

167

Soil parameters were modified by changing parameters in

SPROFILE.PN~,
the soil characteristicsinput file to PNUTGRO. Thus,
no modifications were needed to the PNUTGRO code itself. The Pascal program SENPOS modifies soil parameters as interpreted data
from soil water sensors dictate. The soil input file contains soil characteristics data for 29 different soils. M e n it is run, PNUTGRO automatically picks the proper data set for the soil chosen.
Two specialists initially identified parameters in the soil characteristics file that may be different from actual values for the Millhopper soil used in the study (Smajstrla, 1989) and those that would
influence zone water contents significantly(Hoogenboom,1989).
Based on these discussions,myfamiliarity with the processes involved, and several model runs to determine output sensitivity to
changing parameters, evidence was strong that the DUL (drained upper limit) needed adjustment. The DUL is the water content below
which drainage no longer occurs out of a soil layer. A DUL of 0.110
cm3/cm3was felt to better represent sands in the area (including the
Millhopper soil) than the value of 0.086 cm3/cm3present in the data
file for a ~ i l ~ o p p soil.
e r (This value is close to 0.107 cm3/cm3,the
value for a generic medium sand in the soil data file.)
Although not significant in the high (wet) ranges of modeled
water content, the parameter describing the lower limit of water content (LL) can influence modeled water contents in drier ranges that
occur most often in the shallow soil zones. For this reason, LL and
DUL were adjusted together to approach the values for a medium
sand. The discussions that follow focuson DUL adjustment although
the adjustment method is the same for LL.
Rules were designed and implemented inthe
procedure
SprofileModif to adjust parameters (see Table 1).The body of the
procedure is executed only if the soil is drying. Decisions to adjust
parameters are based on a comparison of model- and sensor-based
representations of the composite tension in the water regulation zone.
Composite tension from the model is determined by averaging the
water contents (cm3/cm3)in each layer within the calculated regulation zone and converting the result to tension (Wa) usinga published
soil water characteristic curve for a Millhopper sand (Carlisle et al.,
1985). Sensor-based composite tension is determined the same way,
although no conversion was necessary because tension was read directly. Only morning readings of soil water tension were used to allow forsoil water redistribution inthe evening (Thomson and
Threadgill, 1987).

le 1 Knowledge System Rules in SENPOS

Procedure: SprofileModif
Rule
IF (composite sensor suction) > (composite modeled suction +
THEN subtract value from DUL in soil characteristics input file.
Rule
IF (composite modeled suction) (composite sensor suction +
THEN add value to DUL in soil characteristics input file.
R e ~ u ~ In
~ sboth
:
cases,DUL is modified by a factor proportional to the
difference in modeled and sensor suction. User is informed of the modification.
Procedure: RootDepth
Rule
IF (one day increase in sensor reading >5)
THEN assign RootDepthSensor to sensor depth.
Rule
IF (one day increase in modeled suction >2)
THEN assign RootDepthModel to depth of sensor in modeled zone.
R e ~ u ~ Consequences
~s:
are used in procedure WRModif to modifyroot
weighting factors. User is informed of rooting depths indicated by both the
model and the sensor.
Procedure: WModif
Rule
IF (RootDepthSensor) > (RootDepth~odel)
THEN subtract values from root weighting factors below sensor depth.
Rule
IF (RootDepthSensor)
(RootDepthModel)
THEN add value to root weighting factors below sensor depth inmodeled
zone.
R e ~ u ~Root
~ s : weighting factors are modified by a value proportional to the
relative magnitude of root weighting factor.

If the model-based representation of composite tension was

greater (or less) by 10 Wa than the composite sensor tension on any
portion of the drying cycle, a small water amount proportional to the
difference between the two tensions was added to (or subtracted
from) the DUL value in SPROFILE.PN2. This is represented by the
equation
C =: 0.O21(Sm - S,)/AF

where

C = amount of water added to or subtracted from DUL (cm3/

cm3)
S, = model-derived composite tension value in root zone (ma)
S, = composite sensor tension in root zone (Wa)
AF = adjustment factor (or "gain" term) = 100
The constant (0.021) in the equation is the difference between the
desired DUL (0.107) and the DUL in the soils file (0.086).
Using a small tension difference (10 Wa) as the criterion for DUL
modification allowed parameter adjustments to occur in the early
days of drying so the chance to modify the DUL would not be missed
because of a possible rainfall event. It was noticed during the field
study that extended periods of no rain were rare. If a larger tension
difference were used as criterion for changing the DUL, the entire
adjustment of 0.021 could conceivably be made all at once. However,
a large tension difference might take several days to occur since tension increases nonlinearly (more slowly under wet conditions) due to
the soil characteristic curve. Thus, if a rainfall event were to occur
during that period, no adjustment could be made. Small increments
were added to or subtracted from the DUL because a small observed
tension difference (10 Wa) provided relatively weak evidence that
the DUL needed adjustment.
2.

In PNUTGRQ, root water uptake is a strong linear function of root

length .density and contains a very small logarithmic component.
Root length density, in turn, is weighted by a root length density
factor that is a linear function of the root weighting factor WR(I).
WR(I), therefore, can be used to modulate root water uptake. In determining WR(I),PNUTGRQ accounts for the genetic capability of
different crops or cultivars to proliferate in the root zone. The WR(1)
function was derived for a sandy soil from experimental data of Robertson et al. (1980) and is specified for many fine sands in the soil
characteristics input file SP~QFILE.PN2.As the WR(1) function significantly influences the degree of modeled water uptake in the soil
zones, reliable estimates of the function are crucial.
The root function derived as the starting point (before modification) was a composite of the Robertson et al. (1980) curve and an
extrapolation (Fig. 4). Examination of soil water sensor data at the
lysimeter indicated that water uptake observed using sensors oc-

0.9

.-.-

0.8
0.7

Jonesand Kiniry (1 986)

Robertson et al. (1 9801
Extrapolatedusing sensor data

0.6
-4

0.5
0.4
0.3
0.2
0.1

0
0

20
80 60 40

100

120 160
140

(cm)
FIGURE4 Root weighting function WR(1) as a function of depth.

curred at much shallower depths than the original curve indicated.

The extrapolated curve was anapproximation but wasfelt to be more
consistent with observed sensor responses at the lysimeter installation. Another study by Boote et al. (1989b) alsoindicated a shallower
rooting system for peanuts under similar conditions.
Two sets of rules were needed to modify the root weighting
factors. A rule was needed to infer where water use was occurring
based on sensor readings. A similar rule inferred where water use
was occurring from the model. A resulting depth of water use was
determined at the output of each rule. Based on a comparison of
where water use was occurring, another set of rules modified the root
weighting factor. Table 1lists the rules described below in procedures
RootDepth and WRModif.
The procedure RootDepth looped through all individual sensors
in the field starting withthe shallowest to determine if water uptake
had been detected. If readings over 1 day increased more than 5 Wa,
the variable RootDepthSensor was set to the corresponding sensor
depth. Similar criteriawere used for model outputs. If model-derived
tension values increased more than 2 Wa over a 1 day period, the
corresponding depth was assigned to the variable RootDepth~odel.
A greater tolerance was used for the sensor because readings in the
field can fluctuate slightly. Human error in readin the sensor and
temperature effects exposing small inaccuracies in sensor ca~bration

rocess Simulatio~

171

equations derived for this study (Thomson and Armstrong, 1987) are
potential sources of error.
If the sensor-determined rooting depthwas unequal to the
model-determined rooting depth, the procedure WRModif was activated. The procedure modified root weighting factors so that future
model-based representations of water use distributions in sensor
zones could converge on sensor-based representations. The equation
to modify the root weighting factors was
Wr,,,

= Wr ir F(Wr/Wr~*~)

where
Wr,,, = new root weighting factor
Wr = previous root weighting factor
F = adjustment factor (or "gain" term) = 0.075
Wrtot= sum of root weighting factors in root zone
As with modification of the DUL, the gain term was chosen to provide a conservative adjustment. This could be done because the root
weighting function was reasonablyclose to the "sensor-inferred"
function initially.

Routines in the program ask the user about irrigation management

objectives including the composite tension to trigger the irrigation
system, The user is also asked what percentage of root mass to replenish with irrigation water and what type of irrigation system is
being used. Solid set sprinkler or center pivot systems are the two
choices available. Managem~ntdiffers greatly between the two systems, as center pivots require an early start-up so the crop in the last
part of the field irrigated is not water-stressed.
An important calculation determines the water regulation zone.
The user-entered value of percentage replenished root mass and
model outputs of root length in each zone determine the depth of
this zone. All composite water potentials calculated from the
PNUTGRO crop model used to trigger irrigation are calculated within
this regulation zone. Irrigation is also applied to the bottom of this
zone. The user receives feedback from the program regarding suitability of sensor placements (if enough were placed in the zone) and
is told to what depth the zone will be replenished corresponding to
the entered value of percentage root mass.
A program procedure determines a projected irrigation date and
calculates the amount of water to apply to the field. The projected

172

T~omso~

irrigation date is based on model outputs of soil water content in

each soil zone several days ahead and depends on the type of irrigation system. Soil water contents are converted to tensions using a
published soil water characteristic curve [in this case, for a Millhopper sand (Carlisle etal., 1985)l.A composite tension in theregulation
zone is calculated and compared to the user-entered value. When
the user-entered trigger level is reached or exceeded, irrigation is
recommended. The system assumes no rainfall as worst-case condition. If rainfall occurs, the user-entered rainfall updates the prediction.
The amount of water to applyis based on the difference between
the water content corresponding to the projected tension on that date
and the soil water content at field capacity in the zone of water regulation. The soil water content at field capacity is assigned the value
of drained upper limit (DUL) in the first five soil horizons listed in
the soil characteristics filefor the soil already hard-coded into thesoil
characteristics file. This soil characteristics file,SPROFILE.PN2, is an
input file to PNUTCRO. The DUL can change as the knowledge base
is run, as it is one of the parameters altered by sensor-based feedback.
Procedures were written to provide water amount corrections forthe
next irrigation cycle. The deepest sensor to respond within 1 day of
irrigation indicates the depth of water penetration. If the depth of
penetration is different than the calculated zone of regulation, the
procedure AdjustIrrig adds or subtracts a small amount (2.5 mm) of
water to be applied to the next irrigation cycle. Since this adjusted
amount must be used for later activations of the program, it is stored
to and accessed from a permanent file.
A procedure, DayToIrrigate, provides different ~anagementfor
cases where the system applies water u ~ f o r m l ywith time (as in a
solid set system) or nonuniformly (as in a center pivot system). For
a solid set system, it is assumed that the entire field is irrigated at
once. Thereis, therefore, no lead time required to apply water. Center
pivot systems provide added complexity if a model-based prediction
is used to schedule irrigation. In DayToIrrigate, the user is asked to
enter the amount of water the system applies and thetime to traverse
the entire field at the 100% travel, This information is used to determine the lead time needed to start the system so the crop is not
stressed at the end of the cycle.
To apply the center pivot scheduling algorithm properly, some
assumptions were made. First, the model-based predictions should
represent conditions in the first part of the field irrigated. Likewise,
all soil water sensors should be confined to the first part of the field.

Drying occurs after a saturating rain where field soil moisture conditions are assumed to be uniform. Nonuniform soil moisture conditions occur after the first irrigation cycle, so sensor readings cannot
be averaged if the sensors were widely spaced.
For the present strategy, a primary requirement is that the system arrive at the last part of the field on time so the crop does not
stress. If calculations show that the center pivot cannot apply the
required amount of water in time, a lesser amount is applied proportional to the system capacity and the time required to reach the
end of the field. Theuser is told how much water should be applied,
how much will actually be applied, and the percentage timer setting
required to run the system.

.
To test the parameter adjustment methods of
SENPOS,
daily
PNUTGRO runs using the IREP weather data were made. The purpose of the test was to determine if model-based representations of
soil water potential converged on sensor readings as SENPOS adjusted the three selected model input parameters simultaneously. The
testalsoverified
proper operation of each parameter adjustment
routine.
l.

Drying periods July 26-30 and August 19-23 were selected to illustrate convergence of soil water potentials. To facilitate the test, shortcuts were taken to enter data quickly using a screen editor instead of
running the system in the user mode (illustrated in Fig. 3).
Data from a field test weather file were imported by PNUTGRO.
Soil water sensors began to show drying on July 27, Julian day 207
after a rain event. This was consistent with soil water drying trends
noted in the PNUTGRO output file. PNUTGROsown calculations of
soil water redistributio~after a rain event set the initial conditions.
Root weighting factors were set to values derived from examination
of soil water sensor data for the entire season. This is represented by
the extrapolated curve of Fig. 4 and served as the initial best guess
rooting function.
Soil water sensor data from the north lysimeter for July 27 and
28 were manually entered into a data file input to SENPOS. A weak
drying response was noticed from two shallow sensors placed in the

17

T~omso~

south lysimeter. Thiswas probably due to thesensors not being close

enough to the active roots. Onesensor also indicated a rapid wetting
response compared to the others, indicating that applied water was
probably reaching the sensor through fissures created at installation.
Thus, sensors in the south lysimeter had to be relocated due to improper installation and were not used for the July runs.
SENPOS was then run on July 28 (day 210). Both model- and
sensor-derived values of soil water content and soil water tension in
each sensor zone were recorded. After the July 28 run, PNUTGRO
was run again with any parameters that might have been changed
by SENPOS. During each PNUTGRO consultation, soil characteristics
that are used in the simulation are displayed to the user. The user
can see if any parameters have been changed by SENPOS. For this
study, the parameters DUL and WR(1) in each soilzone were recorded
before each new run of PNUTGRO. SENPOS and PNUTGRO were
run alternately in this manner until July 31.
Other runs were implemented between August 19 and 23 (days
232-236). The firstrun of SENPOS was on August 21 (day 234) using
sensor data for August 20 and 21. Parameters that were adjusted at
the end of the July runs served as a starting point for the August 21
run. The August runs of SENPOS and PNUTGRO were carried out
the same as the July runs except that sensor readings were averaged
for the two lysimeters.
Figure 5 illustrates that the DUL for the top fivesoil zones
(down to 60 cm) increased with each run of SENPOS until day 235.
Composite tension values determined by the model were greater than
composite sensor readings, necessitating a daily change in DUL values. Figure 6 indicates that the model-based tension began to converge on the sensor-based tensionin July and continued convergence
in August. The results of Fig. 6 represent the combined effects of
changing root weighting factors [IWR(I)]as well as the PUL and LL
values in a 24 cm zone of water regulation. For this study, the water
regula~onzone corresponded to about 70% of the root mass (or root
length assuming a linear relationship) as recommended by Richards
and Marsh (1961) and Taylor and Ashcroft (1972). A procedure in
SENPOS determined 28 cm to be the water regulation zone corresponding to 70% of root mass. A 24cm zone was calculated in a
SENPOS procedure to be the closest discrete point to 28 cm at the
edge of the 20 cm sensors zone of influence. (A SENPOS procedure
calculated a 12 cm zone of influence around each sensor placed or
20 cm deep.)

75

e l f - A ~ j u s t Process
i~~
I

0.13 I
0.12
0.11
0.10

0.09
0.08

0.07
0.06

0.05

0.04
0.03

0.02

0.01
0
213212211210

237 236"
235 234

FIGURE 5 Drained upper limit 0)UL) value of the top fivesoil zones as
modified by parameter adjustment routine.

Figure 7 illustrates modifications to the root weighting factors,

WR(I), in July (days 210-214) and August (days 235-238). The July
plot (Fig. 7a) shows that weighting factors were reduced on day 212
in the 15-30 cm zone and were progressively reduced below 30 cm
during the entire period. Progressive reductions below 30 cm did not
influence water use calculations in the 24 cm zone of regulation because reductions in the lower zones were not added proportionately
to the upper zones. The August plot (Fig. '7%) shows that WR(1) increased for day 235 in zones below 30 cm. The factors were reduced
at shallower depths in August than any reductions that occurred in
July. WR(1) reductions in this zone would influence model-basedwater use calculations in the 24 cm regulation zone.
Figure 8 illustrates how sensitive the output tensions (converted
from PNUTGRQ water content using the characteristic curve) are to
changes in the root functions. The soil DUL and LL were set to 0.107
cm3/cm3 and 0.035 cm3/cm3,respectively (the desired convergence
values), to create these graphs. The modified rootfunction yields tension values that are closer to sensor representations in both cases
before the adjustment algorithms further modify the function (as
shown in Fig. 7). Of course, the user may not have the luxury of
appro~matingthe root function with field data. If the user were to

600

500

Model, Daily Modified Parameters

"
+

Sensors

.-

300

[CT

200

L.

-.-0

100

0
208

209

210

21 1

21 2

Model, Daily Modified Parameters

IGURE 6 Composite soil water tension results from sensors; model using
unmodified DUL, LL, and VVR(1); and model using daily modified DUL, LL,
and WR(1) for (a) July 26-30 and (b) August 19-23 in 24 cm zone of
regulation.

0.9
0.8

0.7

--

0.6

0.4
0.3

0.2
0.1

0
210

21 1

212
Julian Day

2;3

214

0.9
0.8

0.7
0.6
0.5

0.3
0.2
0.1

0
235

237

236

238

Julian Day

0-5cm

5-1 5 cm

15-30cm

30-45cm

45-60cm

60-90cm

IGURE 7 Root weighting factors in several soil zones modified by the parameter adjustment routine for
July and @) August.

use the empirical function described by Jones and

first appro~mation,converg
sented in Fig. 6 could still be
established in a short time.
if the user were to run the
model with the rooting function derived from Robertsonet al.
appropriate rules would be needed to alter the overall ~ n c t i o nrap-

208

209

2io

2i1

2i2

Julian Day
FIGURE8 Comparison of modeled soil water tension using three root functions to sensor tension at
8 cm depth and (b) 20 cm depth.

idly. Rapid adjustments such as this could cause oscillation as the

root function attempts to close in on the correct value. In the absence of other rooting data for peanuts, it is suggested that a user
running this knowledge system (or PNUTGRO alone) use the modified root function described herein. For other models following the
IBSNAT (1986) format, the empirical function described by Jones and
n i y (1986) should be used as a first approximation in the absence
of reliable crop- or soil-specific data.

Test of l r r i ~ a t i o ~

Sc~e~~ler

The irrigation scheduling algorithms were tested using data at the

same lysimeter site that was used to evaluate the parameter adjustment algorithms. The lysimeters were irrigated using sensor readings
as these algorithms had yet to be finalized. To test the irrigationscheduling portion of the decision support system, the system was
run using weather data from the lysimeter test and finalized algorithms. The following illustrates decisions made by the system regarding irrigation.
A test of the irrigation-schedu~gportion of the decision support system was performed over one season on irrigated peanuts
grown in twolysimeters. Rainfall, irrigation, soil water sensor potentials, and actual ET were measured. Figures 9-12 show s o i l water
tension data from sensors for the growing season at the 8,20, and 40
cm depths, respectively. The first major drying response that could
indicate pending irrigation began on Julian day 210 as seen in the
figures for the north lysimeter. A composite tension value for these
three depths was calculated by the program. The composite sensor
value was not great enough to warrant irrigation, and the lysimeters
were not irrigated, as illustrated in Fig. 12b.
Although irrigation did not take place (Fig. 12),a model run on
day 210 indicated that irrigation was needed because its calculated
composite tensionvalue was greater than a user-entered trigger value
of
@a. The model indicated that 1.31 cm of water was needed to
replenish a water regulation zone of20 cm. This 20cm regulation
zone was also calculated in the program. At this point, model-based
representations of soil water tension were beginning to converge on
sensor readings as a key s o i l parameter (drained upper limit or DUL)
and root weighting factors were being adjusted, but convergence was
not established until the next significant drying cycle (occurring in
August). During July, irrigation was called for on days 210 and 212.
Thus, the conservative parameter adjustments implemented could
not permit the model to be used reliably by itself this early in the
season.Based on convergence occurring in August,however, the
model could be used reliably once it had been "trained' by the sensors during the "second round' of adjustments.
Drying cycles of a reasonable duration (necessary to activate the
correction algorithms) were few and far between as indicated by a
lengthy period (20 days) between the July drying cycle and the August drying cycle. As has been mentioned, convergence was related
to the degree of adjustment allowed (or a "gain" term). Conservative

130

100
90
70

.g

60
50
40
30
20

0
Julian Day

l80
2 0
290 270 250 230 210 ld0

220

240

260

280

300

Julian Day
ICURE 9 Soil water tension data at 8 cm depth for Watermark sensors in
(a) north lysimeter and (b) south lysimeter.

S e ~ f - A ~ j u s tProcess
in~
Si~ulation

(a)

Julian Day

FIGURE10 Soil water tension data at 20 cm depth for Watermark sensors

in (a) north lysimeter and (b) south lysimeter.

T~omson

34
32 30 28 26 24 "Tii 22
f?i. 2018
16 -c

-:

E ;;r

10 8642-

I 8 0 I 200
230210190

(a)

220

240

1 260 I 280
250 290270

300

300

Julian Day

28 I
26 24 22 20 18 a, 1
25 6'
.-g 14
12108m

::

6 -

42 ' l80

210
190

200

220 I 240
250
230

260

290
270

2bO

Julian Day

11
water tension data at 40 cm depth for Watermark sensors
north lysimeter and (b) south lysimeter.

IGURE

in

Self-A~jutin~
Process Simu~ation

183

180 I
170

1
130
120

E
c
L

1 1 0 ~
100 t"

g
2

$j

90
80
70
60

50
40
30
20
10
0

(a)

Julian Day
12
11 10

9-

8 76 -

5 4 -

3211

27 0

280

Juiian Day
JGURE 12 Water applications (a) from rainfall and (b) from supplemental
irrigation.

adjustments have to be made under conditions like these when evidence is not very strong that a particular parameter needs adjustment. ~ u n n i n gthe system in a drier region could yield many more
drying cycles, allowing full corrections to occur early. This would
allow even more conservative adjustments than were already effected. A system that would work better under drier conditions is
desirable because d v regions are, of course, more critical with regard
to irrigation. For this study; it is fortunate that adjustments were
made because model-based predictions left uncorrected would have
aled many unnecessary irrigations.
Environmental conditions and characteristics of the soil at the
lysimeter site provided a stringent test of the method as evidenced
by the next irrigation event at day268. On this day; 0.931 cm of water
was applied to the lysimeters based on sensor readings and a composite trigger level of 60 kPa (Fig. 112b). There was some question as
to whether irrigation should be delayed because of forecast rainfall.
For this reason, it was decided that the amount of water applied
should be a little less than the amount needed to replenish a full 40
cm water regulation zone. (The sensors indicated significant water
uptake at the 40 cm depth as seen in Fig, 11.)Therefore, the amount
of water applied was calculated to be the amount needed to refill a
cm zone to field capacity(or the drained upper limit). This amount
could allow for additional recharge by rainfall. Irrigation was decided
on because sensor tension levels were climbing very rapidly due to
the low water retention characteristics of the illh hopper sand. Indeed, enough rain to replenish the zone fell within a couple of days
after irrigation commenced.
A run of the model also indicated that irrigation was warranted
on day268. This late in the season, parameters had long since settled
on their "static" values, and model-based soil water tension was apking the sensors quite well. However, the amount of wawas calculated to be 2.5 cm, an amount required to replenish a zone of 40 cm. The model does not presently account for
prevailing weather conditions and thus does not allow a deficit irrigation strategy. Theuser would need to override the calculated water
amount if less water were desired.

Although present parameter adjustment routines adjust only soil water parameters and the rooting function, evapotranspiration (ET)
should be considered because inaccuracies in modeling crop water

use can also influenceresults. Many trends can be observed by comparing model-based ET with measured ET.
Figure 13a compares simulated and measured ET rates over the
season. By examining trends over a shorter time scale, more detailed
observations can be made. Figure 13b illustrates an ET comparison
over the period August 12-29 (days 225-242). Data for days 231237 represent a drying cycle immediately after a high intensity rain,
and ET measured during this period should be close to potential ET
rates. During this period, measured ET was greater than simulated
ET by an average of 28%. This might be partially explained by sensible heat advection occurring over exposed lysimeters. The quantity
LE/ Rn (latent heat of vaporization of water times ET rate divided by
net radiation) can be used to ascertain sensible heat advection (Verma
and Rosenberg, 1977). If LE/ h > 1.0, sensible heat consumption is
occurring at the soil surface. Over a sunny period of days 230-237,
LE/Rn averaged 1.47. It showed a low of 1.09 on day 230 and a peak
of 1.85 on day 236. Measured ET was significantly higher than simulated ET during periods of sensible heat consumption.
All days where simulated ET was greater than measured ET
were characterized by low radiation loads (with or without rain). On
days 226 and 238, for example, weather was cloudy with no precipitation. PNUTGRO used a form of the Priestley and Taylor (1972) ET
model, which included an advection modification based on air temperature (Ritchie, 1985; Jones and Kiniry, 1986). Ritchie (1985) used a
multiplier of 1.1 on the ET equation to account for the effects of unsaturated air and increased the multiplier to allowfor advection
when the maximum temperature was greater than 24C. In PNUT6110,however, the multiplier was increased when the maximum
temperature exceeded 34C. Air temperatures on days 226 and 238
were well below this threshold, so potential ET on these days was
simulated by the unmodified Priestley-Taylor equation. A study by
Steiner et al. (1991) indicated that the unmodified Priestley-Taylor
model under predicted ET at low ET rates and overpredicted ET during periods of high evaporative demand. The authors suggested that
a more responsive advection function could improve predictions at
both ends of the ET spectrum.
If, in the future, rules are developed to modify the ET model,
methods should be applied separately to evaporation and transpiration components. This would be very important in the early stages
of crop development because the evaporation component is a large
part of the total ET. Although sensor-based feedback would be difficult to implement in the earliest parts of the season, information

T~ornso~

Measured
Modeled
"
l

8
a,

U:

.-0

a
v)

--_.

Measured
Modeled

225
227
229
231
233
235
237
239
241

IGURE 13 Comparison of measured and modeled ET rates (a) over peanut

growing season and (b) from August 12 to August 29.

could soon be obtained as roots begin to proliferate. Soil water sensors cannot provide reliable information on soil evaporation in the
shallowest soil horizons, but they can provide information on the
combined evaporation and transpiration in horizons where roots
proliferate.

A sensor-based knowledge system was developed to adjust soil water

and rooting parameters of a crop model. The objective of this study
was to provide a better representation of temporal soil water conditions for better water and chemical management and irrigation scheduling. Less reliance on sensors could be achieved as the model makes
better estimates. An expert system to evaluate the suitability of sensor
readings used to calibrate PNLJTGRO was coded in a commercial
expert system shell, VP-Expert. Pascal routines adjusted drained upper limit (DUL),lower limit of water content (LL), and root weighting
factors [WR(I)] that were input values to the PNUTGRO crop model.
Sets of field data from July and August 2988 verified that modelbased representations of temporal soil water status converged on
sensor-based representations as parameters reached their new static
values. It is envisioned that the approaches and procedures outlined
in this chapter can be applied widely to systems that might benefit
from sensor feedback to calibrate model input parameters that are
difficult to characterize. Other potential applications are discussed in
the next section.
Expert system rules that determined whether a sensor reading
was valid compared a sensors reading with readings from other sensors placed at the same depth. This did not consider the chance of
the entire system drifting. It is conceivable that the entire system
could be off if all sensors drift significantly off calibration, for example. Although significant drift was not noticed over a year of field
installation, periodic calibration checks are recommended.
An unexpected sensor response to wetting was noticed during
the lysimeter study resulting from a rain event. Figure 14 illustrates
that one sensor
lysimeter 2) showed complete rewetting, although several others at the same depth showed no such response.
In this case, there was evidence that the sensor might have been
placed improperly, creating a fissure that established a direct line for
water to saturate the sensor. Although cases like this may be infrequent, decision rules should be developed to handle them.

Thornso

1
170

263

261

259

257

Julian Day
IGURE

14 Sensor responses due to a 0.4

rain event (at day 263).

Additional rules could be developed to elaborate on the characteristics of different sensor types. For example, my experience with
Watermark sensors indicates that the sensors themselves usually do
not "fail'' in the wet range of operation unless wires break or fray. If
one of these sensors does not show water use from drying and other
sensors at the same depth do indicate water use, for example, it is
much more likely that the sensor needs to be relocated to the zone
of active roots. Thesesensors show a very gradual degradation in the
wet range of operation characterized by readings that do not "return
to zero" when saturated (Pogue,1991). Additional decision rules
could be written to accommodate these cases.
It was also mentioned that tensiometers can "break tension" in
the dry ranges of operation (approximately 80 Wa). As coded, there
are no rules that presently accommodate this situation. Breaking tension would be an infrequent problem when scheduling irrigation in
a field with sandy soil. Usually; a crop is never allowed to approach
this stress level because a sandy soil holds very little water at this
tension level. Itis thus common practiceto irrigate much sooner than
80 Wa. By contrast, data from a field study undercorn in Shenandoah
County; Virginia (Fig. 15) showed deep rooting and tension levels
frequently exceeding 80 Wa in some horizons before irrigation was
called for. Figure 15aindicates when tensiometers broke tension and

rocess Si~ulation
80

Julian Day

I
Julian Day

15 Soil water sensor responses under corn in Shenandoah County,

(June 1993). (a) Tensiometer responses;(b) Watermark sensor responses.
= 15 cm depth,
+ = 30 cm depth, and V, 0 = 60 cm depth.)

URE

required recharging. The soil in Shenandoah County was a Wolfgap

loam. Although the 15 cm reading would be disregarded later in the
season, uptake was pronounced at the 60 cm depth quite early in the
season. This contrasts with past observations under corn in a Lakeland sand (Tift County Georgia), for example. Under these conditions, I have not seen significant water uptake (as registered by sensors) at the 60 cm depth (Thomson, 1983). This was not seen to be a
hard pan problem. Corn in the sandy soil required frequent irrigation,
and roots were thus most prominent in the shallow 20-45 cm depths.
Figure 15b is a plot of Watermark sensor readings in the same
row. These readings were taken with the Watermark 30KTC meter
and converted to resistance using an equation found in Thomson
(1990). Readings were further compensated for temperature and converted to tensions (up to 100 kPa only) using the relationship of
Thomson and Armstrong (1987). It is interesting to note temporal
differences between tensiometer and Watermark readings as shown
in the figures. Tensiometers seemed more responsive than Watermark
sensors to wetting and drying. A major drawback, however,was that
they required periodic service as indicated on the graphs. For this
reason, the farm manager taking the data was biased in favor of the
Watermark sensor (Wilkins, 1993). Watermark sensors at the 30 and
60 cm depths seemed rather unresponsive this early in the season
(June), which, initially might indicate improper placement. Later in
the season, however, sensors at these locations began to register good
drying. These comparisons show clearly that further study is needed
to better quantify sensor response characteristics. A well-tuned expert
system could then do a better job interpreting these responses.
Based on observations noted, expert system rules for sensor selection might also be desirable. It is clear from the data of Fig. 15, for
example, that one type of sensor might be better suited than another
for reasons already mentioned. Another sensor that has not been
mentioned is the gypsum block (Delmhorst, 1993). This
sensor works
on the same resistance principle as Watermark sensors and gives
readings up to 1500 Wa tension. Although conventional crop management would not allow this stress level to be approached, resolution is said to be good from about 30 to 150 kPa, an acceptable management range for crops grown in medium and fine textured soils.
These sensors need to be replaced yearly because they dissolve, but
they are the least expensive of the three sensor types mentioned. The
Watermark sensors can be reused, but removing resistance-type sensors from dense, compacted soils can be an arduous task. For this

rocess S i m ~ ~ a t i o ~

reason, a farm manager may opt to leave them in the soil and install
new ones the next year.
PNUTGRO was modified to accept daily ET as input from the
field study so differences in sensor- and model-based results could
be attributed to different soil water and rooting characteristics only.
For actual application, ET cannot be an input, and analysis presented
in Thomson (1990) and herein indicates that the model tracked measured ET very well. Forthis study, modeled ET could have been used
with little influence on the results.
Care must be taken when adjusting parameters that influence
the same variable. The developer should be keenly familiar with the
processes involved and how variables interact. Many model runs and
sensitivity studies can increase the developer's knowledge and thus
allow intelligent examination of variable interactions. For this study
conservative adjustments to parameters were made to achieve the
desired goal. By analogy with control system theory, the output reflected somewhat of an "overdamped" response. This type of response was desired because, as indicated for the DUL, evidence was
not strong enough to warrant large adjustments. In general, a conservative approach is probably the best course to follow until the
knowledge base grows and more evidence (extracted and interpreted
from field data) is presented. I believe that the system was quite stable with the adjustment factors used.
A key factor in acceptance of this method is how universally
applicable it may be. Soil and rooting characteristics, for example,
vary widely with the soil and crop. Running this system on fields
with a few more soil-crop combinations would further strengthen
the knowledge base. Other soil characteristics that were not thought
to be significanthere might require consideration. A major advantage
is that once a robust knowledge base is constructed it can be selfcalibrating. That is,many exhaustive field trials should not be needed
because the rules should "know" what trends to look for. Machine
learning could be used to ensure that the proper parameter was adjusted. If, for example, an inappropriate adjustment was made, later
trends in data might give a clue that the adjustment was wrong. If
an adjustment to the correct parameter were then made and subsequent water status data showed more consistent tracking, the system
could recall this and apply the correct adjustment later if similar environmental conditions were encountered. Self-learning systems that
use neural networks might be used for the parameter adjustment
portion of this work and have shown promise in estimating model
outputs with limited input data (Altendorf et al., 1992).

Except for early in the season, the system worked well as an

irrigation scheduler under the conditions tested. Early in the season,
parameter adjustments had just begun and the model indicated drier
conditions than the sensors indicated. For this reason, the model
called for irrigation when none was warranted. To effect parameter
adjustments, this system requires drying cycles of sufficient duration
that comparisons can be made between sensor tensions and modelbased tensions. Based on the magnitude of the difference, selected
parameters are adjusted with the objective of correcting the simulated
soil water status. As has been stated, there were few drying cycles of
cient magnitude at the lysimeter plots. By contrast, sensor readings during the summer of 1993 at Shenandoah County, Virginia indicated many drying cycles. Drying cycles of sufficient duration occurring early would allow parameters to adjust and stabilize early to
their final static values. Thus, this system should work better under
a drier climate.

As has been indicated in Thomson et al. (1993), this methodology

could find its use in many applications. One application could be in
grain or peanut drying. The following is an untried concept that
could apply the methods of the irrigation problem to the grain drying
problem.
In grain drying, especially in slower, more efficient in-bin drying, there is a problem in the measurement of moisture content of the
grain at different levels in the grain mass, which may be as deep as
10 or 12 ft (3 or 4 m). Grain drying models have been developed that
can accurately estimate the grain moisture content througho~tthe
grain mass, given an initial grain moisture reading, airflow rate,temperature, and dew point of the entering airflow.The initial grain
moisture can bemeasured accurately as the grain is placed in thebin,
and the inputair conditions can be measured accurately overtime as
the drying progresses. Airflow is difficult to measure accurately, however, and inaddition it is felt in the industry that airflow maychange
as the grain shrinks during drying. So the question is: Can an initial
estimate of airflow rate be adjusted during drying by an automatic
expert system, as in the irrigation problem, so that grain moisture can
be accurately estimated by a grain drying simulator? The requirements for solving this problem in a way analogous to the irrigation
problem would require a method of sensing the grain moisture content throughout the bin, and this is manually difficult, as it is nec-

essary to probe 1-4 m into wet grain. However, a similar approach

could be taken using temperature along the depth of the grain as the
measured parameter, as a thin probe containing many thermocouples
for temperature measurement would not be difficult to install.
Then the drying simulation, using the initial estimate of airflow
rate, could be run, and the temperatures at several depths would be
compared with themeasured values. Careful analysis of this problem
could result in rules adjusting the assumed airflow rate to bring the
simulated temperatures closer to the measured temperatures as the
drying continues. Since the conditions of the ambient air going into
the fan and bin are dynamically changing, the heat and mass transfer
modeling in the drying simulation must be validated before confidence can be placed in this adjustment technique. However, several
grain drying models exist, and this technique could allow the use of
simulation to predict the actual drying situation in a bin. This would
allow adjustments, such as turning the airflow off during periods
when no drying is occurring, to be made by an expert system to
reduce cost and improve the quality of the product.
The idea of using an expert system to implement the concept of
adaptive control or machine self-learning can
be realized by methods
given in this chapter. Simulations of processes such as crop growth,
evapotranspiration, and grain drying based on physical and biological principles can be extremely useful, especiallyif they can betuned
or adjusted to fit the particular environment in which they are being
used. These environments, which change with location, include the
water-holding properties of the soil in the irrigation problem and
the properties of the grain mass that affect airflow in the grain drying problem. It is hoped that these principles can be applied by
other researchers to improve the usefulness of agricultural system
simulations.

Altendorf, C. T., M. L. Stone, R. L. Elliot, and M. A. Kizer. 1992.Determining

soil moisture using soil thermal properties. Paper No. 92-3020. St. Joseph,
MI: American Societyof Agricultural Engineers.
Boote, K. J., J. W. Jones, G. Hoogenboom, G. G. Wilkerson,and S. S. Jagtap.
1989a.I'NUTGROV1.02.IBSNATVersion.
User's Guide. Gainesville, FL:
University of Florida.
Boote, K. J.,J. M. Bennett, J. W. Jones, and H. E. Jowers. 198913. On-farm
testing of peanut and soybean models in north Florida. Paper No. 89-4040.
St. Joseph, MI: American/ Canadian Societies of Agricultural Engineers.

19

Thornson

Butts, C. L. 1988. Modeling the evaporation and temperature distribution of

a soil profile. Ph.D. Dissertation, Agricultural Engineering Department,
University of Florida, Gainesville.
Carlisle, V.W.,M. E. Collins, F. Sodek 111, and L. C. Hammond. 1985. Characterization Data for Selected Florida Soils. Soil Science Report No. 85-1.
Gainesville, FL: University of Florida, Institute of Food and Agricultural
Sciences, Soil Science Department Laboratory, p. 169.
Delmhorst. 1993. Application literature for gypsum soil blocks. Towaco, NJ:
Delmhorst Instrument Company.
Hoogenboom, G. 1989. Personal communication. Gainesville, FL: Department of Agricultural Engineering, University of Florida.
IBSNAT.1986. Decision S ~ p ~ o System
rt
for A ~ o t e c h ~ Transfer
o ~ o ~ (DSSA").
Documentation for IBSNAT crop model input and outputfiles. Version1.0.
Technical Report 5.
Honolulu, HI: University of Hawaii, Collegeof Tropical
Agriculture and Human Resources.
Irrometer. 1989. Application literature for Watermark Model 200 soil moisture sensor. Riverside, CA: Irrometer Co.
Jones, C. A., and J. R. Kiniry (Eds.).1986. C E R E S - ~ ~AZ S~~: ~ ~ ~ fModel
ffion
of Maize Growth and ~evezopmen~.
College Station,TX: Texas A&M University Press.
Jones, W., K. J. Boote, S. S. Jagtap, G. Hoogenboom, and G. G. Wilkerson.
1988. SOYGRO V5.41. IBSNAT Version.User's Guide. Gainesville, FL: University of Florida.
Pogue, W. 1991. Personal communication. Riverside,CA: Irrometer Company
Priestley, C. H. B., and R. J. Taylor. 1972. On the assessment of surface heat
and evaporation using large-scale parameters. ~ o n t h Z y W e a t Rev.
~ e r 100(2):
81-92.
Richards, S. J., and A. W. Marsh. 1961. Irrigation based on soil suction measurements. Soil Sci. Soc. Am. Proc. 25(1):65-69.
Ritchie, J. T. 1985. A user oriented model of the soil water balance in wheat.
In: ea^ Growth and ~ o ~ e Z i n E.
g . Fry and T. K. Atkin (Eds.), NATO-AS1
Series. New York Plenum, pp. 293-305.
Robertson, W. K., L. C. Hammond, J. T. Johnson, and K. J. Boote. 1980.Effects
of plant-water stress on root distribution of corn, soybeans, and peanuts
in a sandy soil. Agron. J. 72:548--550.
Ross, B. B. 1984. Irrigation scheduling with a farm computer. Microcomputer
conference and trade show,Mar.19-21.Blacksburg,
VA: Virginia Polytechnic Institute and State University,pp. 23-32.
Smajstrla, A. G. 1985. Design and management of drip irrigation systems for
tomatoes. Agric. Eng. Extension Mimeo Report
85-13. Gainesville, FL: Florida Cooperative Extension Service,IFAS, University of Florida.
Smajstrla, A. G. 1989. Personal communication. Gainesville, FL: Department
of Agricultural Engineering, University of Florida.
Steiner, J. L., T. A. Howell, and A. D. Schneider. 1991. Lysimetric evaluation
of daily potential e~apotranspirationmodels for grain sorghum. Agron. J
83(1):240-247.

ocess

Self-Adjusting

195

Taylor, S. A., and G.L. Ashcroft.1972. Physical E d a ~ h o l o ~

The
: Physics of
Irrigated and on-Irrigated Soils. San Francisco, CA:W. H. Freeman.
Thomasson, J. A. 1993. Foreign matter effects on cotton color measurement:
determination and correction. Trans. ASAE 36(3):663-669.
Thomson, S. J. 1983.Unpublished tensiometer and Watermark response data;
water management under center pivot irrigation. Tifton, GA: Coastal Plain
Experiment Station.
Thomson, J. 1990. Knowledge system for determining soil water status
using sensor feedback. Ph.D. dissertation, University of Florida, Agricultural Engineering Dept., Gainesville, FL. Ann Arbor, MI: University Microfilms, Intl.
Thomson, S. J., and C. F. Armstrong. 1987.Calibration of the Watermark
Model 200 soil moisture. Appl. Eng. Agric. 3(2):186-289.
Thomson, S. J., and E. D. Threadgill. 1987. Microcomputer control for soil
moisture based scheduling of center pivot irrigation systems. C o ~ p ~Elect.
tron. A g ~ i c 1(4):321-338.
.
Thomson, S. J., J. W. Mishoe, R. M. Peart, and F. S. Zazueta. 1989, IEEE-488
Interface and Instruments for High ResolutionMeasurement of Data.
Paper No. 89-3560. St. Joseph, MI: AmericanSociety of Agricultural
Engineers.
Thomson, S. J., R. M. Peart, and J. W. Mishoe. 1993. Parameter adjustment
to a crop model using a sensor-based decision support system. Trans. ASAE
36(1):205-213.
Verma, S. B., and N. J. Rosenberg. 1977. The Brown-Rosenberg resistance
model of crop evapotrans~iration-modified tests in an irrigated sorghum
field. Agron. 1, 69:332-335.
Wilkerson, G.G., J. W. Jones, K. J. Boote, K.
Ingram, and J. W. Mishoe.
1983. Modeling soybean growth for crop management. Trans. ASAE 26:
63-73.
Wilkins,C.1993.
Personal communication.Wilkinsfarm,
Shenandoah
County, VA.

This Page Intentionally Left Blank

Although the evaporation of water from land surfaces and from vegetation is a process of profound importance for agriculture, ecology,
and geophysics, mechanismsgoverning the process remained obscure
until relatively late in the history of science. In 1867, G. J. Symons
described evaporation as the most desperate art of the desperate
science of meteorology, and it was an art that made little progress
until well into the 20th century.
Reviewing the partitioning of available energy at the earths surface, Brunt (1939) wrote: A fraction, whose ~ a ~ i t u isd not
e ~eadil~
e s t i ~ a t e d[italics mine],is used in evaporating water from the surface
of grass, leaves, etc.and inhigh latitudes a portion is used in melting
ice and snow. A few years later, however, the first measurements of
fluxes of water vapor over uniform vegetation by Thornthwaite and
Holzman (1939) signaled the start of major advances in both experimental and theoretical aspects of evaporation science. Progress was
then interrupted by World War 11, and it was not until 1948 that the
publication of three seminal papers initiated a field of research that
is still expanding 50 years later.
197

onteit

At Rothamsted, England, Penman (1948) used a set of small lysimeters, installed many years previously, to measure evaporation
from open water, bare soil, and grass. By combining a thermodynamic equation for surface heat balance and anaerodynamic equation
for vapor transfer, he derived a general formula for the rate of evaporation from open water as a function climatic elements (temperature, vapor pressure, wind, and radiation) measured at screen
height. The key to this analysis was the elimination of surface temperature (a rarely measured and often unmeasurable quantity) which
was achieved by assuming that saturation vapor pressure was a hear function of temperature over small ranges. If Penmans equation
had been published a few decades later it would almost certainly
have been known today as Penmans model!
Working independently in the USSR, Budyko (1948) used the
same primary equations as Penman but estimated surface temperature by a process of iteration, thereby deriving estimates of evaporation more accurately than Penman but more laboriously. (This1948
monograph was an English translation of the original Russian text
that predated Penmans classic paper.)
Thornthwaite (1948) proposed a completely differentsolution to
the problem of estimating evaporation from watersheds in the United
States, given minimal climate records. Like Penman, he argued that
evaporation was driven primarily by energy available fromradiation,
but, for the practical reason that very few measurements of radiation
were available, he used air temperature as a surrogate variable. This
argument is still valid in some circumstances, and a recent prediction
of global evaporation by Mintz and Walker(1993) used Thornthwaites method in preference to Penmans.
This review of evaporation model opens with a brief description
of Thornthwaites formula but deals mainly with developments from
the theoretical base established by Budyko and Penman. One major
concept emerging from their work was the distinction between potential rates of evaporation from vegetation or soil obtained when
the supply of water was abundantand actual rates prevailing
when the supply was restricted.

After examining a large number of hydrological records for climatically contrasting sites in the United States, Thornthwaite concluded
that the potential evaporation in any month could be correlated with

199

~va~oration
Models

the mean monthly temperature (C) by using a scaling factor 1 defined as

where is mean monthly temperature and the summation is over 12
months. The scaling factor was then used to define an exponent a
such that

E = 16(10T/I)

(2)

where

a = (0.67513 - 77.11

+ 17,9201 + 492,390) lo6

This set of equations contains eight numerical constants, all determined empirically and some quoted with unwarranted precision. As
needed, the constants were modified to allow for the dependence of
daylength on latitude and for monthly temperatures outside the
range 0-26.5C.
In a well-balanced critique of Thornthwaites formula, Pelton et
al. (1960) concluded that it was most reliable for periods of a month
or longer because, over a year and at many sites, monthly mean temperature and potential transpiration were strongly correlated. However,
Because of the severe limitations of mean temperature methods, they will yield correct s ~ o r t - ~ e r estimates
io~
of evapotranspiration only under fortuitous circumstances and cannot be
relied on for general use . . . To expend effort improving mean
temperature methods for short-period potential evapotranspiration estimates and adjusting the estimates to a given
locale and crop when mean temperature has little or no physical foundation for this use appears to be kicking a dead
horse.
Despite this unequivocal warning, agronomists and engineers,
especially in the United States, continued to refine and use empirical
models for estimating potential evaporation from mean temperature
(Blaney and Criddle, 1962) or diurnal temperature range (Hargreaves
and Samani, 1982). When locally calibrated, they perform well considering how little input they need as shown in a recent comparison
by Allen et al. (1994). However, becausethey are products of current
climate they are prone to error when used to predict changes of evaporation rate as a consequence of global warming, particularly in dry
climates as shown by McKenny and Rosenberg(1993).Empirical

o~teit

models based on solar radiation (e.g.,Makkink, 1957; Jensen and

Haise, 1963) also perform well in environments where they have been
calibrated.

Starting from the basic physics of heat and vapor transfer, Penman
(1948) developed a mechanistic model for evaporation from free water surfaces and extended it empirically to bare soil and grass. This
model was a practical application of the First Law of Thermodynamics which requires that the net rate at which radiant energy is absorbed at any point on the earths surface, often written R, (W/m2),
must be balanced by the rate at which energy is (l)transformed to
latent heat of evaporation, XE (W/ m), where h (J/ g) is the latent heat
of evaporation for water and E [g/ (m. S)] is a mass flux of water
vapor, or (2) dispersed to the atmosphere by convection, or (3) stored
in a substrate (soil, rock, water, etc.). (Energystored by photosynthesis, essential for growth, is energetically negligible.)
Because instruments for measuring R, were not available in the
1940s (and are still very rare at climatological stations 50 years later),
Penman calculated this quantity from its components. Incoming
short-wave (solar) radiation was estimated from hours of sunshine
using a relation derived by Angstrom [cited by Brunt (1939)] and net
long-wave radiation was estimated from air temperature and vapor
pressure using an empirical formula derived by Brunt.
Further empiricism was needed to obtain (1) a vapor transfer
coefficient (the flux of water vapor per unit difference in vapor pressure between a water surface and the air passing over it at some
arbitrary height) and (2) a corresponding heat transfer coefficient. For
both heat and vapor, the transfer coefficient used by Penman was
obtained from measurements of evaporation from small (76 cm diameter) evaporation tanks and had the general formf(u) = a ( l + bu),
where a and b were empirical constants and u was wind speed measured ata standard height of 2 m. This function was used to estimate
fluxes of both sensible heat and latent heat from a surface, given air
temperature and vapor pressure (as measured at the height of a climatological screen) and the egective temperature and vapor pressure
of a wet surface.
By eliminating surface temperature and vapor pressure from this
set of equations, Penman was able to express the latent heat of evaporation from a wet surface as
XE = (AH + YE,)/ (A

where A is the rate of change of saturation vapor pressure with temperature (which increases rapidly with temperature from 45 to 145
Pa/K between 0 and 2OOC) and y is the psychrometer constant in the
same units of pressure per unit temperature difference. (The factthat
this constant is a weak function of both temperature and pressure
is usually ignored. At a standard pressure of 101.3 Wa, it increases
from 65 to
Pa/K between 0 and 20C.
In Penmans 1948 paper, H was the amount of energy per unit
of ground surface as received from radiation, less a fraction stored in
the soil. (A few workers, including the writer, still observe this tradition, but most use the symbol H for sensible heat flux.) Penman
assumed that over periods of a week or longer, soil heat
storage could
be neglected compared with the net amount of heat received from
radiation.
Using contemporary units, the composite variable E, can be
defined as
E a

PADftU)

(5)

where p is air density (g/m3),A @ / g )is the latent heat of vaporization

of water, D (kg/kg) is the mean specific humidity deficit of air at
screen height, and f(u)(m/ S) is a linear function of wind speed at a
standard height, usually 2 m.
Dividing the numerator and denominator of Eq. (4) by y and
writing E = A/y gives

Penmans model combined robust thermodynamic and aerodynamic principles and provided a sound basis for theoretical developments in evaporation science. Itcould not be used in practice, however (to estimate irrigation need, for example), without recourse to a
substantial amount of empiricism.WhereasThornthwaites model
contained eight empirically determined constants, Penmans, in its
original form contained six-four for estimating net radiation from
The water vapor content of air is conveniently expressedas dimensionless specific
humidity q (kg of water vapor per kg of moist air), related to vapor pressure by
=
[ ( p - e)
==
where = 0.622 is the ratio of molecular weights for
water vapor and air, e is vapor pressure, and p is air pressure in the same units.
The approximation is valid because e is invariably two orders
of magnitude less
than p . The parameters A and y then assume units of K- and y = c,/X (K-), where
cp[J/ (kg-K)] is the specific heat of air at constant pressure.

202

~onteit

sunshine hours and vapor pressure and two for the wind function.
A seventh constant was needed to estimate the "actual" rate of evaporation from soil or vegetation as a fraction of the "potential" rate
for open water in the same environment. Subsequent theoretical and
instrumental developments, now reviewed, made much of this empiricism unnecessary.

A major generalization of Penman's original equation was made possible by introducing "resistances" -simple electrical analogs for the
potential difference needed to drive unit flux in systems that involve
the transport of momentum, heat, and water vapor (Monteith and
Unsworth, 1990). (Such resistances have dimensions of time per unit
length and are usually quoted in units of seconds per meter or per
centimeter.)
What subsequently became known as the en man-Monteith
equation (PME hereafter), incorporating both an aerodynamic resistance Y , and a surface or stomatal resistance Y,, was first published in
a somewhat obscure contract report from a group led by F. A. Brooks
at Davis, California (~onteith,1963), and a more extended treatment
followed (Monteith, 1965). An almost identical contemporary model
was developed wholly independently by Rijtema (1965).
In the PME, the empirical wind function f(u) used in Penman's
equation was replaced by the reciprocal of an aerodynamic resistance
so that the second term in the numerator became

E: =

(7)

and the psychrometer constant was modified to

Y" = Y(Ys + ra)/ ra

so that the PME could be written in Penman's format as
XE = (AH + YE:)/ (A

+ 7")

(9)

The appearance of a wind-dependent function in the denominator as well as in the numerator of Eq. (9) implies that the rate of
evaporation calculated from the PME is always less dependent on
wind speed than the rate from the corresponding Penman equation
when other elements of climate are unchanged. In general, estimated
rates are usually insensitive to the magnitude of r,, and the error

203

generated by neglecting the influence of buoyancy correction (see

later) is often small. In contrast, evaporation rate is usually a strong
function of surface resistance (Fig. 1).
Ry differentiating Eq.
with respect to l/ra, it can be shown.
that d/d(l/ ra) = 0 when the surface resistance is
Y,

= pc,(A-'

+ y")D/H

(10)

and substitution of this value of surface resistance into Eq.

X/H = A/(A

gives

+ 7)

It follows that when the surface resistance is close to the value set by
Eq. (lo), the evaporation will be insensitive to wind speed and close
to AH/(A + y) (Fig. l; see also Section IV).

90

IGURE
Dependence of latent heat of evaporation hE on aerodynamic
resistance r, and surface resistance r, when total available heat H = 105
W/m2, saturation vapor pressure deficit D = 0.2 @a, and A = 0.15 W a / K
In mostcircumstances, evaporation rate decreases as Y, increases (top
curve). However, there is critical value of Y, (90 s/m in this case) at which
h E / H = A/(A +
[see Eq. (15)] so that E is independent of
When r, is
smaller than this value, the evaporation rate increases as ra increases.

When the concept of a "surface resistance" was first introduced,

it was severely criticizedby a group of Australian meteorologists who
were convinced that the complexprocesses of heat and vapor
exchange in a canopy could not and should not be modeled in such
a simplistic way [see Appendix to Monteith (1963)l. Equation (9)has
stood the test of time, however, and has been used on many scales
to estimate evaporation from singleleaves, from crop stands, and
from extensive areas of uniform vegetation in models of the general
circulation of the atmosphere.
Sources of Error

Because the net heat term AH is usually several times larger than AE,,
it is the main source of error in the numerator of the PME. This error
has three components: the net flux of radiation received by vegetation
over the period of application; the net amount of heat stored in the
vegetation; and the net amount of heat stored in the soil. Forfractions
of a day or for forests with a large heat capacity, the magnitude of
the storage term must be evaluated and retained if it is comparable
with the net radiative flux over the same period [see Thom (1975),p.
93 et seq.].
Net radiation can either be measured directly or estimated from
its components. Unfortunately, net radiometers are notoriously prone
to several types of error, mainly associated with changes in the transmission properties of polythene domes during exposure (Field et al.,
1992; Duchon and Wilk, 1994). Climatological measurements of net
radiation are therefore both rare and unreliable. Records of solar radiation are more trustworthy and more readily available but must be
complemented by estimates of long-wave incoming radiation (usually
obtained from empirical formulas) and from outgoing radiation assumed equal to blackbody radiation at air temperature because the
radiative surface temperature is rarely available. However,the difference between air and surface temperatures can be accounted for by
defining an effective radiative resistance as rR = pcP/4aT':, where a is
Stefan's constant and T',is the air temperature in Kelvin (Monteith
and Unsworth, 1990). Thevalue of this resistance decreases from245
s / m at 10C to 186 s / m at 30C.If the effective resistance for heat
transfer by convection and radiation, treated as parallel processes, is
taken as (ri' + ri')-', it is legitimate to estimate net radiation as the
flux density Xni (W/mz)for a surface at air temperature Ta ("C) (sometimes referred to as the "isothermal" net radiation). For clear skies
an appropriate algorithm for the long-wave component of net radiation .Rnl (Monteith and Unsworth, 1990) is
Rd
(12)
= 107 - 0.3Ta

c.
Within comprehensive models of crop growth and yield, the PME
can readily be exploited as a submodel for transpiration either at the
scale of a single leaf or at canopy level provided the relevant physical
and physiological variables are specified a priori or are available as
output from other parts of the model. Following an exhaustive comparison of lysimeter measurements of evaporation and predictions
from a range of formulas, Allen et al. (1994) concluded that the PME
was more reliable than other models over a wide range of climates.
The PME has therefore been adopted by FAO as a substitute for the
Penman equation modified by Roorenbos and Pruitt (1977) in an irrigation manual usually referred to as FAO-24. Several other evaporation models, including the original Penman equation, fitted Allen's
measurements data almost as well as the PME and can therefore be
confidently used for practicalapplications when the surface resistance
is known.
Doorenbos and Pruitt altered Penman's original formula by introducing an alternative wind function that predicts a smaller value
of the aerodynamic resistance at a given wind speed. However, Thom
and Oliver (197'7) demonstrated that an error in the original wind
function (which was appropriate for a smooth water surface rather
than for rough vegetation) fortuitously compensated for the assumption that the surface resistancewas effectively zero. Eventually, Allen
et al. (1994), using measurements from high-class lysimeters, demonstrated conclusively that when this compensation was removed,
the "improved' formula overestimated evaporation by 10-30% (the
precise figure depending on the relative magnitude of radiation and
humidity deficit terms).
Since 1981, the British Meteorological Office has used the PME
operationally over the whole of Britain to provide weekly and
monthly estimates of evaporation and soil water deficit (Thompson
et al., 1981). The correct type of wind function was used along with
empirical relations between surface resistance, leaf area (Grant,1975),
and soil water potential (Russell, 1980).

It has been shown that both aerodynamic and surface resistancesplay

an important role in establishing rates of evaporation from land surfaces.Their properties are now examined morecloselyforsingle
leaves, for uniform stands of vegetation, and for bare soil.

206

Single Leaves

For the simplest case of a single transpiring leaf, the total resistance
to the diffusion of water vapor from substomatal cavities to the external airstream is analogous to an electric current passing through
two resistances in series, usually referred to as "stomatal" and "aerodynamic" resistances.
The magnitude of the stomatal resistance can be estimated in
principle from the number of stomata per unit leaf area and from the
mean diameter and length of pores. These are the main variables of
diffusion models for stomata (e.g., Penman and Milthorpe, 1967), but
because they are rarely known and are difficult to determine with
precision, stomatal resistances are usually calculated from measured
rates of transpiration and estimated gradients of vapor concentration.
When stomata are closed,leaves of most species continue to lose
water very slowly by diffusion through waxy cuticles, but their resistance is so large compared with stomatal resistance that it is usually assumed to be infinite,
There is an aerodynamic resistance r, to diffusion in the boundary layer surrounding the leaf within which the transfer of heat, water
vapor, etc. proceeds at a rate governed by molecular diffusion. Provided the wind speed is great enough and the temperature difference
between leaf and air is small enough to ensure that transfer processes
are not affected by gradients of air density, the boundary layer resistance depends on air velocity and on the size, shape, and attitude of
the leaf with respect to the airstream. In very light wind, however,
rates of transfer are determined mainly by gradients of temperature
and therefore of density, so that the aerodynamic resistance depends
more on the mean leaf-air temperature difference than on wind
speed. Empirical formulas for estimating leaf boundary later resistance as a function of wind speed and characteristic dimension can
be obtained from the engineering literature as presented by Monteith
and Unsworth (1990) and by Jones (1992), for example.
To obtain the correct total (stomatal plus aerodynamic) resistance to vapor diffusion, rst, it is necessary to distinguish between
"amphistomatous" leaves with stomata on both abaxial (ab) and adaxial (ad) surfaces and "hypostomatous" leaves with stomata on adaxial surfaces only. Forthe amphistomatous class, the total resistance
is found by combining the resistance for the two surfaces in parallel
to give
For the h~ostomatousclass, the resistance is simply Y ,

+ r,.

Models

207

Models for the rate of evaporation from a single leaf (or of an

individual plant treated as an assembly of leaves) have been used by
glasshouse engineers seeking an optimum thermal environment for
growth or production by taking account of the balance between the
market value of a crop and production costs, including heating or
cooling. Such models are usually based on engineering literature for
heat transfer from small plates of regular shape (e.g., circular or rectangular). For example, Seginer (1984) suggested that the resistance
to heat transfer by free convection from rose leaflets (both surfaces)
in virtually still air could be expressed (in s/m) as 660,' (6T/d)0.25,a
quantity obtained from an appropriate Grashof number (as given in
the ASHRAE handbook, for example). The quantity 6T is the mean
temperature difference (K) between the surface of the leaf and the air
surrounding it (as measured with an infrared thermometer, for example) and d is a characteristic leaf dimension.
The corresponding resistance to water vapor transfer raV depends, in principle, on the ratio of molecular diffusion coefficients for
heat and the gas in question, but the difference between resistances
for heat and water vapor is usually small enough to neglect.
The same procedures are valid for single leaves outdoors with
the obvious complication that wind speed and direction are usually
very variable. In this case, the aerodynamic resistance may be estimated from a Nusselt number depending on the characteristic dimension of the leaf in the average direction of the wind and the
component of average wind speed parallel to the surface ( ~ o n t e i t h
and Unsworth, 1990).

In the field, aerodynamic resistances for homogeneous surfaces such

as bare soil or a crop canopy are large-scale analogs of the boundary
layer resistances for individual leaves and other organs. In principle,
resistances for heat, water vapor, and other scalars can be estimated
as the gradient needed to maintain unit flux, but in practice they are
often derived from measurements of wind speed and from a knowledge of the aerodynamic properties of the surface, using the following
procedure.
When the atmosphere above an extensive uniform stand of vegetation is in a state of neutral stability (i.e., when the temperature
gradient is close to the dry adiabatic lapse rate or - l0C/ 100 m), the
increase of wind speed U with height z above the ground is given by

The so-called friction velocityu*(m/s) is defined by

where T is the momentum flux from atmosphere to surface (often

referred to as a shearing stress) and is air density. Von Karmans
constant C
I, as determined experimentally, was assumed for many
years to be 0.41, but recent measurements suggest that 0.397 0.010
maybe a more appropriate value [ see,e.g.,Frenzen
and Vogel
(1995)l.
The ability of a surface such as the canopy of a crop to absorb
momentum (at a given wind speed) is specified by a roughness
length zomeasured not from the ground surface but from the efective
height of the canopy as defined by a zero plane displacement d. For
many years zo and d were estimated empirically as fractions of crop
height h (e.g., zo = O.lh, d = 0.7h). A model for momentum transfer
within canopies developed by Shaw and Pereira (1982) provides a
more mechanistic basis forestimating zoand d, and an application of
this model by Choudhuryand Monteith (1988) was successfully
tested by van den Hurk et al. (1995). The main structural parameter
of the model is the product X of leaf area index L, and the mean drag
coefficient of individual leaves c& Zero plane displacement and
roughness length are then given by
d = l.lh(1

+ Xo.25)

(15)

and
or
zo = 0.3h(l

d/h)

when

0.2 S X

1.5

(16b)

and z; is the roughness length for the soil surface.

The resistance to m o m e n t u ~transfer raMbetween the surface
and a height z can be obtained from Eqs. (16a) and (16b) because
it is the momentum differenceacross the resistance or p[(u(z) u(z - d)] = pu(z) divided by the momentum flux pu:. This implies
that

so that raMis inversely propo~ionalto u(z) provided other parameters

are independent of wind speed. In practice, because most vegetation
has leaves, petioles, or stems that bend in the wind, d tends to decrease with increasing wind speed [as shown by Vogt and Jaeger

(1990), for example], whereas zo may either decrease or increase depending on changes in foliage density. Businger(1956) suggested that
Tanner and Pelton (1960) confirmed that the reciprocal of raMis equivalent to the wind function f(u) in Eq. (5).
When foliageis warmer or cooler than the air above it, mechanically generated turbulence is respectively enhanced or inhibited by
the action of buoyancy. For these so-called non-neutral conditions,
Eq. (17a) can, in principle, be modified by introducing a stability parameter such as the Richardson number or the M o n i n - ~ b u ~ o v
length (Thorn, 1975; Monteith and Unsworth, 1990), but the former
requires measurements of temperature and wind speed gradients that
are not generally available and the latter requires a sensible heat flux
that cannot be obtained without iteration.
An aerodynamic resistance for the transfer of scalars such as
heat or water vapor can be obtained by adding to the momentum
resistance raMan additional component rb to account for the absence
of a process equivalent to the blufiody forces that enhance momentum transfer ( ~ o n t e i t hand Unsworth, 1990). For arable crops and
with U* in the range 0.1-0.5 m/s, Thorn (1975) suggested the empirical relation rb = 6 . 2 ~ , . ~Alternatively,
~.
heat and vapor resistances
can be obtained directly from q. (17a) by replacing zo with scalarecific roughness lengths SUC as zoE3for heat or zov for vapor in
place of the omentum roughness usually written zo. Equation (17a)
can then be rewritten for heat or water vapor as

Garratt (1992) suggested that ln(zo/zoII) ln(zo/zov) 2 was appropriate for diverse types of vegetation, but from the analysis of many
measurements over short grass, Duynkerke (1992) reported values of
ln(z,/zoE.3)
r a n ~ n gfrom about -2 to 13 and increasing both with leaf
area index and with U*. Precise modeling of evaporation rates therefore requires a careful assessment of differences in the flux- adient
relation for heat, for vapor, and for momentum.
Because it is possible to identify an aerodynamic resistance for
extensive uniform vegetation [using Eq. (17a)], it is also possible to
identify a surface resistance imposed by stomata, assuming that a
complete canopy is endowed with the properties of a single big
leaf. As for individual leaves, the surface resistance of a canopy
is usually estimated indirectly from fluxesand gradients using a procedure discussed in the next section.

n v e ~ s i oof
onteith
~ the

E ~ ~ ~ t i o n

Whereas the original Penman model has been used mainly to estimate the irrigation need of crops or seasonal changes in the water
balance of catchments, a common use of the PME is to estimate the
surface resistance of vegetation when its rate of ~anspirationhas been
measured independently The magnitude of this resistance can then
be used as anindex of ground cover, water stress, nutrient deficiency
etc. The procedure is to invert Eq. (9) to obtain

where variables on the right-hand side of the equation are measured

or estimated. In some of the first published measurements of surface
resistance (Monteith, 1963), r, for irrigated grass at Davis, California,
achieved minimum values in the range 0.5-1 s/cm in the middle of
the day. Morning values were somewhat larger, and late afternoon
values were substantially larger. Such behavior suggested that the
value of r, as evaluated from Eq. (18) encapsulated the dependence
of stomatal aperture on weather and water supply Many subsequent
measurements supported this conclusion and confirmed thatthe
range of resistances observed at Davis was characteristic of many
types of vegetation [as cited by Kelliher et al. (1995), for example].

As a basis for estimating surface resistance in prescribed weather or

interpreting measurements in terms of weather variables, Jarvis
(1976) introduced the concept of a maximum conductance gm(the
reciprocal of a minimum resistance). Thisconductance is achieved in
principle (but rarely if ever, in practice) at simultaneously optimal
values of solar radiation S incident on a canopy; air temperature Ta
and vapor pressure deficit L) both measured at screen height;and soil
water content averaged over the root zone (0). The actual conductance
of a leaf at any time is given by

where each of the component functions fi, fi, etc. has a m a ~ m u m

value of 1. The general form of these functions is known from laboratory measurements on single leaves (e.g., fi is commonly a rectangular hyperbola or negative exponential function),but the parameters
defining the functions have to be determined experimentallyby using
Eq. (19) overas wide a range of environmental conditions as possible.

Evaporation Models

Equation (19) has been used to relate the effective stomatal/

surface resistance of canopies to variables measured above the canopy notwithstanding the fact that stomata respond to the temperature
ofleaf tissue rather than to air temperature, to the vapor pressure
deficit at the leaf surface rather than at some undefined distance from
foliage, and tothe whole range of solar irradiance experienced ~ithin
a canopy rather than the irradiance measured above it. These anomalies can be partly but not completely circumvented by treating foliage as a system with n discrete horizontal layers, each with a leaf
area index of Z(n) and a conductance g(n) depending on local illumination and temperature, age, etc. For the simplest case of layers in
parallel, the bulk conductance is Zg(n)Z(n)provided the aerodynamic
conductance between layers is much greater than g(n) (McNaughton
and Jarvis, 1983). This is rarely a realistic model, however, so conductance is usually estimated by inverting the PME applied to the
whole canopy as already discussed.
A second type of stomatal model developed by Ball et al.(1987)
was based on the thesis that stomatal aperture is tightly coupled to
the net rate of photosynthesis as determined by biochemical processes
that are functions of irradiance, temperature, etc. In this treatment,
stomatal conductance of single leaves is expressed as

where h, and c, are respectively the relative humidity and carbon

dioxide concentration at the leaf surface, A is the net rate of photosynthesis, go is the cuticular conductance of the leaf as measured
when stomata are shut so that A is virtually zero, and k is a constant
with appropriate dimensions.
Though based on careful laboratory work and much used, this
model is mechanistically unsound because there is no evidence that
stomata respond to relative humidity as distinct from absolute humidity a quantity that is often an acceptable surrogate for transpiration rate.This anomaly was resolved by Leuning (1995), who
showed that laboratory measurements of photosynthesis and stomatal conductance were consistent with an equation proposed by
Lohammar et al.
namely

where D, is the specific humidity deficit at the leaf surface and Dsois
a scaling factor. As the transpiration rate E is given by gD, (assuming

21

o~teit

negligible temperature gradient across the cuticle), the stomatal component of resistance is given by

For the common condition that g is much larger than go, rearrangement of Eq. (22) gives

g = g& where g, (-kj4/cs) is the apparent maximum conductance when E: =

0 and E , (-g,r>,,)
is an apparent maximum transpiration rate
achieved when stomata are closed. Reanalysis of many laboratory
measurements (Monteith, 1995) confirmedthe general validity of Eq.
(23) within the experimental range of saturation deficit and Bunce
(1997) has suggested that peristomatal transpiration is implicated.
Notable applications of Eq. (19) were analyses of evaporation
over pine forest at Thetford in southeast England (Stewart, 1988)and
in the Les Landes region of southwest France (Cash et al. 1989). Although multiple regression analysis applied to both sets of measurements yielded similar sets of functions as shown in Fig. 2, this agreement does not guarantee that all three functions are appro~riate.The
radiation function has the general hyperbolic form observed in many
laboratory experiments. The temperature responseresembles(but
1.Q

0.5

""".
c

Canopy conductance (as fraction of maximum value) for Scots

Pine forest in Thetford, southeast England (---) and Maritime Pine forestin
the Les Landes district of southwest France (-). Plots represent statistical
fits to functions assumed for dependence of conductance on solar radiation,
temperature, and specific humidity deficit. (From Gash et al.,1989).

cannot be compared directly with) laboratory measurements, which

are almost invariably referred to the temperature of foliage rather
than thatof the ambient air. Thehumidity response implies that when
saturation deficit increases from zero, the evaporation rate increases
as stomatal conductance decreases, and this behavior has been observed in the laboratory, Above a saturation deficit of l0 g/kg, however, the surface resistance remains fixed as the evaporation rate continues to increase in response to an increasing deficit. As this is
inconsistent with all the laboratory evidence on which Eq. (23) is
based, it is likely that the humidity response incorporates a fallacious
artifact.
Other examples of deriving stomatal functions from field measurements can be found in the work of Dolman (1993), who used the
Jarvis function to analyze measurements over Amazonian forest, and
of Dougherty et al. (1994), who fitted measurements of stomatal conductance on C4 grasses to the original Ball-Berry model and five
other superior models derived from it. In the best model,conductance
was shown to be proportional to a power of the function A/Dc,,
which has several features in common with Eq. (21).
The substantial amount of scatter observed when measured Values of surface conductance are correlated with values calculated from
a set of regression functions for individual parameters has several
potential sources: errors of observation, weakness of the assumption
that normalizing functions can be treated as independent multipliers
[Eq. (19)1, neglect of heat and vapor storage in canopies, and the
adoption of a physiologically unsound function for the putative response to saturation deficit.
If, as demonstrated in the studies cited, the response of the apparent surface resistance orconductance of canopies to environmental
factors is similar to the response of single leaves, some measure of
correlation would be expected between maximum values of conductance observed on single leaves in the laboratory (gmax) and m a ~ m u m
values of canopy conductance observed in the field (Gmax).Kelliher et
al. (1995) demonstrated that such congruence does exist, assisted by
the fact that the rate of evaporation used to estimate G,, often includes a component of soil evaporation. This decreases as leaf area
index increases in such a way that G,, is less dependent on leaf area
than might be expected. To summarize conclusions fromthe Kelliher
et al. study, mean values of gmax
and G,, ranged from 6 to 20 mm/
S for natural vegetation and from 12 to 23 mm/s for a~icultural
crops. The ratio G,,,/g,,,
was about 3 1 (see Fig. 3).

onteith
I

~ - - r

FIGURE3 Maximum canopy conductance G,,,, for a range of vegetation

types plotted against maximum leaf conductance g,,,,, with (bold) line of
best fit through
Thin lines are model estimates of the relation for leaf
area indices of 2 and 10. Key: tun, moss and tundra/lichen; gra, grassland;
cer, wheat; bhc, arable crops; wsh, heather and macchia shrub; euc, eucalyptus forest; dec, deciduous forest; con, coniferous forest; and trf, -tropical forest. (From Kelliher et al., 1995.)

The model used above to explore the exchange of heat and water
between vegetation and the atmosphere was essentially an electrical
analog containing a network
resistances but neglecting capacitances-elements of the system capable of storing heat or water for
periods usually longer than an hour but less than a day. Their properties are now considered briefly.
1 . Heat Storage i n

and Vegetation

Diurnal changes in the storage heat in the soil are needed for precise hourly estimates of evaporation using Penman-type equations.

21 5

They can be accounted for in a somewhat crude way by making the

heat flux at the soil/air interface a sinusoidal function of time as in
the ~houdhury-Monteith (1988) model. However, van den Hurk and
McNaughton (1995) showed that this leads to substantial errors that
are avoided when the "force-restore" method of Deardorff (1978) is
used to estimate soil heat flux.
Storage of heat in vegetation is usually negligible for arable
crops but can be a significant component of the heat balance in forests. Following Thom (1975), the maximum rate of change of mean
biomass temperature may beassumed equal to the change of air temperature aTa/at (Klh), and the specific heat of vegetation may be
taken as 70% of the value for water, i.e., 2.9 kJ/ (kg*K).Then for a
biomass of m kg/m2, the maximum change of heat storage is equivalent to a flux of about 0.8 m a?l,/dt. This term will always be negligible for short vegetation with m 10 g/m2.But for a representative
change in temperature of 3 K/h, the corresponding flux for a forest
with m 10 &/m2 is 24 W/m2, a significant quantity with the same
order of magnitude as fluxes of sensible heat.
2.

Water Storage

Passioura (1983) suggested that when the roots of a uniform crop start
to extract water at a specified depth, the volumetric water content 0
can often be expressed as a negative exponential function of time,
e.g.,
0 = 0, exp(--t/.r)

(24)

where 0, is the water "available" at t = 0 and I is the time constant

of the system analogous to the time constant CR of an electric circuit
contai~ng
a capacitance C and a resistance R.
A time constant for roots at a specific depth can be estimated
by measuring the local water content of soil using a neutron probe,
for example, provided the profile is not rewetted by rain. When 0,
was estimated from the difference between water held at field capacity and permanent wilting point (as determined from laboratory samples), values of I for a stand of sorghum growing on a vertisol at
yderabad (India) ranged from 35 to 40 days compared with 10 days
for millet on an alfisol (Monteith, 1984). Subsequently, when Robertson et al.(1993) and Singh et al. (in press) used the difference between
maximum and minimum values of 0 as observed in the field instead
of arbitrary conventional limits for "available" water, they obtained
values of I in the range 15-20 days.

Models of extraction of water by plant roots are still at a primitive stage of development compared with models for the evaporation
of extracted water. When the rate of evaporation is determined by
weather, this disparity is usually tolerable, but when growth is limited by the availability of water to roots, most models resort to somewhat crude empiricism and predictions are therefore unre~able.
3.

WaterStorage Within Plants

The concept of capacity is also relevant to the flow of water through

plants because cellsin stems and leaves contribute to the loss of water
by evaporation during the day and are recharged at night. This process must be taken into account in any detailed model of the time
course of water loss by plant stands, but few relevant measurements
or models are available. Hunt et al.(1991) provided detailed estimates
of hydraulic resistances and capacities for roots, stems,and leaves of
different types of vegetation and computed corresponding time constants from the product of resistance and capacitance. These range
from a minimum of about 20 S for grass with a capacitance of 1 m3/
Pa to about 5 h for trees with a capacitance of 1000 m3/MPa.
WaterStorage Within Canopies

hen rain starts to fall on vegetation, leaves and other surfaces intercept a fraction that depends on factors such as area, angular distribution, and roughness. Eventually the rate of interce~tionis balanced by the rate of loss by dripping from wet surfaces, and the
a ~ o ~ ofn water
t
retained after saturation is known as the canopy
storage capacity (S), a quantity usually lying in the range 0.5-2.0 mm.
Rutter et al. (1975), Calder (1977, 1986), and others explored the relation between interception and evaporation rates by writing the net
~ )a balance between (l)
rate of change of intercepted water ( ~ I / das
the rate at which water is intercepted, assumed to be a fraction of
the pre~ipitationrate P, and (2) the rate at which it is subsequently
lost either by evaporation ( E ) or by dripping [assumed to be a function f(I) of I ] .It follows that

Rutter et al. (1975) assumed that f(I) was a positive exponential

nction of I and that E was equivalent to the potential rate E, obt a i ~ e dfrom the Penman-Monteith equation multip~edby I / S when
I was less than S. Calder used E , in all conditions and showed that
measurements in a spruce forest appeared to fit the simple linear

model f(I) = bI. The value of b depended on the choice of parameters

in the Penman-Monteith equation used to estimate E, and also
showed a weak dependence on the value of I.

The Penman equation implies that when energy is supplied at a rate

H to a wet surface exposed to saturated air ( D = 0), then the latent
heat of evaporation is

and is therefore independent of wind speed [see Eq. (ll)].More generally provided the saturation deficit of air is constant with height
so that air in contact with foliage has the same value ( D ) as air at
screen height, the latent heat of evaporation is

When this equation is used to eliminate D from Eqs. (8) and (IO),
rearrangement of terms again gives Eq. (26).
In practice, saturation vapor pressure deficit is not constant with
height. In the absence of cloud and for reasons outlined below, it
increases with height within a convective boundary layer (CBL) that
is often several hundred meters deep at dawn, expanding to several
thousand meters in depth as the day progresses.
The CBL is capped by an inversion through which relatively
warm dry air is mixed into the boundary layer from above. At the
ground, however, air passing over vegetation receives water vapor
by transpiration and may be heated or cooled depending on the difference between surface and air temperature. Commonly the net flux
of water vapor into the CBL is small, so vapor pressure changes little
during the day. In contrast, temperature invariably increases in response to radiative heating and to an input of sensible heat from
above the CBL that is modulated but never reversed by sensible heat
exchange at the ground. An important consequence of these intimately linked processes is that saturation vapor pressure deficit increases with height in such a way that rates of transpiration exceed
predictions from Eq. (26) by 20-30%. To deal with this excess empirically Priestley and Taylor (1972) suggested that the equation should
be modified by introducing a multiplier (a)such that

21

o~teit

Using this Priestley-Taylor equation (PTE) as a definition of a,the

PME, Eq. (9),can be rewritten in terms of a by multiplying both sides
of the equation (A + ?)l AN to give
l + N
It-X
where N = pc,D/ANr, is a nondimensional climate number depending on radiation, temperature, vapor pressure, and wind speed; and
a="----

is a nondimensional resistance ratio depending primarily on the

relative size of surface and aerodynamic resistances but with a weak
temperature dependence through the value of A.
Many early attempts to test the PTE appeared to be successful
insofar as values of a close to 1.26 were obtained both for water
surfaces and for diverse types of well-watered vegetation. Eventually,
however, it was demonstrated that a could achieve values as large
as l1 for spruce forest with wet foliage (Shuttleworth and Calder,
1979) and values much smaller than 1.26 over vegetation with a restricted water supply as discussed later (McNaughton and Spriggs,
1989).Considerable caution is therefore needed when Eq. (29) is used
with some arbitrary value of N (usually between 0.2 and 0.3) to estimate a "potential" rate of evaporation. In the CERES family of crop
models, this potential rate is multiplied by a ratio of actual to potential evaporation obtained from the PME. However, the PTE and PME
yield the same estimate of potential evaporation only when the surface is completely decoupled from the atmosphere (see below).
Jury andTanner (1975)drew attention to anomalously large values of a obtained in the presence of advection, e.g., when air from
dry land passed over an irrigated field. They thereforesuggested that
the Priestley-Taylor coeffkient should be expanded to include an
empirical saturation deficit term. This procedure has not been widely
adopted, however, partly because it departs from the simplicity of
Eq. (26) without improving on the physics behind Eq. (S), and partly
because major errors arise when empirical constants derived in one
climate are transferred to another where the rainfall and ~umidity
regime are different.

McNaughton and Jawis(1983) explored the advantages of combining

the PTE and PME by writing

S1 = [l

+A+y

(")l
r,

-1

In a system where r, is very large compared with rs so that S1

1,
the latent heat flux is given by AH/ (A + y), which is the flux obtained
from the PTE with a = 1. In this case, the vegetation is regarded as
effectively "decoupled' from the atmosphere above it, implying that
the saturation deficit at the surface is controlled by physical processes
at the surface. Conversely, when r, is very small compared with r,,
0, the latent heat flux is given by the second term in Eq. (31),
and the saturation deficit is imposed on the system by the state of
the air mass passing over it.
This type of model has helped to explore and interpret the contrasting behavior of evaporation from vegetation with relatively
smooth surfaces such as short grass (decoupled from the air passing
over it) and from very rough surfaces such as forests (tightly coupled). In particular, it has thrown light on the significance of evaporation from water intercepted by foliage. The relativelysmall amount
of water intercepted by short vegetation evaporates very slowly because the total rate of evaporation plus transpiration is controlled by
availableenergy.Forests intercept much larger amounts of water,
which evaporate rapidly because of rapid mixing, and this process is
now considered in more detail.
*

ary Layer

Most models of evaporation incorporate climatic variables such as

temperature and vapor pressure that are assumed to be independent
of the state of the underlying surface. In practice, however,
these variables depend on the way air mass properties determined by weather
on a synoptic scale interact with surface properties on a local scale.
Feedback generated by these interactions can be explored within
models for the CBL.
Complete CBL models include dynamical elements, but McNaughton and Spriggs (1989) showed that a simpler thermodynamic
model ofCBL growth was consistent with one of the best sets of
measurements available from a site in the Netherlands (Driedonks,
1981). They then used a subset of the same measurements to predict

how the daily loss of water by evaporation should change in response

to changes in surface resistance, assumed constant throughout the
day. To present this relation, they plotted the Priestley-Taylor coefficient a as a function of the logarithm of surface resistancer, to obtain
a set of sigmoid curves. However, it is more informative to plot the
reciprocal of a against r, because this yields a set of straight lines (Fig.
4) defined by an intercept a. and a slope m = d(1 /a)/ dr, so that

+ my,

a- =

(33)

For evaporation from an extensive region, the climate number

N (Eq. 29) is expected to increase as surface resistance increases(during a spell of dry weather, for example) because of an increase in

a!

100

surface r e s i s ~ (SI
~c~
Reciprocal of the Priestley-Taylor coefficient, a, estimated as a
function of surface resistance by McNaughton and Spriggs (1989), who coupled a boundary layer model to meteorological measurements for 9 days
over grassland at Cabauw, Netherlands (Driedonks, 1981). Nine lines join
sets of points at seven arbitrary values of surface resistance assumed constant
throughout the day so that each set represents daily mean estimates (a)-1
as
a function of resistance for daylight hours on 1 day. Sets of points for each
day are displaced vertically to avoid confusion. Values of for zero resistance ranged from 1.30 to 1.57 with mean of 1.40 0.1.

vapor pressure deficit. The resistance number R Eq. (30) is a function

of aerodynamic as well as surface resistance.
By eliminating a from Eqs. (29) and (33), it can be shown that
when the surface resistance is r,, the ~ o t e ~ t i arate
2 of evaporation (expressed as a Priestley-Taylor coefficient) is given by

where f = ?/(A + ?>.

From Eq. (33), the actual rate of evaporation is
A =

1 + aornrs

(35)

Alternatively, by eliminating r, from Eqs. (29) and (33), the

Priestley-Taylor coefficient for actual evaporation can be expressed
as
a=

1t-N-X
l + f - a&

where X = f /mr, can be regarded as a climatological variable depending on mean air temperature and wind speed and on the struc-

The rate at which water evaporates from wet soil depends partly on
the state of the atmosphere as specified by radiation, wind speed,
humidity, etc. and partly on the state of the soil as established by the
distribution of water held in pores and by vertical gradients of temperature and vapor pressure. Fully comprehensive models of evaporation from soil, as described by ten Berge (19901, for example, account rigorously for the way mass and heat transfer are coupled
within soil and at the soillair interface.
illel (1976) described a representative model of
soil evaporation in which a loam soil with a depth of 1.13 m was
treated as 14 layers of increasing depth starting with a l cm layer at
the surface. Soil water potential and hydraulic conductivity were
specified functions of water content, and fluxes of water and heat
were assumed to be independent rather than coupled as in the more
rigorous model of Philip and de Vries (1957). The short-wave reflectivity and long-wave emissivity of the soil surface were functions of

onteit

222

o Bingharnton

Phoenix
0

0
0
3

**

**

0.4

0.8

FIGURE5 Rate of evaporation from bare soil estimated from the model of
Van Bavel and Hillel and plotted (as fractions of maximum rate) against the
inverse square root of time fromthe onset of drying. Sites chosen for climatic
contrast were Binghamton, New York, and Phoenix, Arizona. (From Van
Bavel and Hillel, 1976.)

the water content of the top layer, and the aerodynamic resistance
between the surface and a reference height of 2 m was found by using
conventional equations incorporating wind speed, roughness, and a
stability correction.Climate records for seven contrasting stations
within the United States were used to explore the behavior of the
model when run with an hour