This article was downloaded by: [University of Boras]

On: 05 October 2014, At: 06:38

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Geocarto International
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/tgei20

Estimating fractional land cover in

semi-arid central Kalahari: the impact
of mapping method (spectral unmixing
vs. object-based image analysis) and
vegetation morphology
a a
Niti B. Mishra & Kelley A. Crews
a
Department of Geography & the Environment, The University of
Texas at Austin, Austin, TX, USA
Accepted author version posted online: 17 Jan 2014.Published
online: 31 Mar 2014.

To cite this article: Niti B. Mishra & Kelley A. Crews (2014) Estimating fractional land cover
in semi-arid central Kalahari: the impact of mapping method (spectral unmixing vs. object-
based image analysis) and vegetation morphology, Geocarto International, 29:8, 860-877, DOI:
10.1080/10106049.2013.868041

To link to this article: http://dx.doi.org/10.1080/10106049.2013.868041

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
 Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
Downloaded by [University of Boras] at 06:38 05 October 2014
 Geocarto International, 2014
Vol. 29, No. 8, 860–877, http://dx.doi.org/10.1080/10106049.2013.868041

Estimating fractional land cover in semi-arid central Kalahari: the

impact of mapping method (spectral unmixing vs. object-based
image analysis) and vegetation morphology
Niti B. Mishra* and Kelley A. Crews

Department of Geography & the Environment, The University of Texas at Austin, Austin, TX,
USA
(Received 16 June 2013; final version received 18 November 2013)
Downloaded by [University of Boras] at 06:38 05 October 2014

Focusing on the central Kalahari, this study utilized fractional cover of

photosynthetic vegetation ( fPV), non-photosynthetic vegetation ( fNPV) and bare soil
( fBS), derived in situ and estimated from GeoEye-1 imagery using Multiple End-
member Spectral Mixture Analysis (MESMA) and object-based image analysis
(OBIA) to determine superior method for fractional cover estimation and the impact
of vegetation morphology on the estimation accuracy. MESMA mapped fractional
cover by testing endmember models of varying complexity. Based on OBIA, image
was segmented at five segmentation scales followed by classification. MESMA
provided more accurate fractional cover estimates than OBIA. The increasing
segmentation scale in OBIA resulted in a consistent increase in error. Different veg-
etation morphology types showed varied responses to the changing segmentation
scale, reflecting their unique ecology and physiognomy. While areas under woody
cover produced lower error even at coarse segmentation scales, those with
herbaceous cover provided low error only at the fine segmentation scale.
Keywords: semi-arid savanna; fractional cover; MESMA; Object-based image anal-
ysis; GeoEye

1. Introduction
Savannas are geographically extensive ecosystems (nearly 40% of global terrestrial area
and 50% of Africa) that are characterized by the unique coexistence of herbaceous and
woody life forms (Mistry 2000; Hill et al. 2011). In southern Africa, savanna ecosys-
tems are both economically and ecologically significant containing large wildlife habi-
tats and also provide the basis of economic activity by supporting livestock ranching
and wildlife based tourism (Moore & Attwell 1999; Lal 2004; Hill et al. 2011). Over
the last few decades, due to increasing anthropogenic impact (e.g. changing land use,
altered fire regime) coupled with climatic variability, vegetation structural and func-
tional attributes in these xeric systems have been subjected to large-scale changes
(Scholes & Walker 1993; Hill et al. 2011). These changes are in turn affecting the eco-
logical dynamics and availability of habitat-related key structural properties (e.g. animal
migration routes, solitary nesting trees and forging areas) (Dougill et al. 1999; Tews
et al. 2004; Blaum et al. 2007).

*Corresponding author. Email: [email protected]

© 2014 Taylor & Francis

 Geocarto International 861

Effective management of savanna ecosystems requires understanding the spatio-tem-

poral variation in ecological processes. This understanding depends on our ability to
accurately characterize and monitor the structural and functional attributes of the
savanna vegetation (e.g. vegetation cover, density and condition) (Scholes & Walker
1993; Mishra & Crews 2014). In savanna ecosystems, vegetation cover is an important
attribute that determines the ability of the system to capture rainfall. Thus, decreasing
vegetation cover is considered an important indicator of land degradation in savannas
(Ludwig et al. 2007; Okin et al. 2009). Further, in these ecosystems the exchange of
energy and water balance is controlled by transpiration and evaporation through the
proportion of photosynthetically active vegetation ( fPV), non-photosynthetically active
vegetation ( fNPV) and bare soil ( fBS) (Scanlon et al. 2005; Guerschman et al. 2009).
Specially, in the semi-arid savannas, fNPV is intimately related to fire frequency and
intensity (Roy et al. 2011) and fBS controls wind and water erosion (Okin et al. 2009).
Field-based measurements of vegetation properties in savannas are limited in scope and
Downloaded by [University of Boras] at 06:38 05 October 2014

scale, especially when considering systems that are vast, remote and wild. Remote
sensing provides tool for estimating fractional cover of vegetation as an indicator that
can not only complement field measurements but also provide much larger spatial
coverage (Elmore et al. 2000; Asner et al. 2011).
For estimating fractional land cover, linear spectral mixture analysis (SMA) is a
popular technique that provides sub-pixel abundance estimates based on the assumption
that the spectral signature of a pixel is a linear, proportion-weighted combination of
endmembers (i.e. pure spectra of ground components) (Roberts et al. 1993; Adams et al.
1995; Guerschman et al. 2009). While the accuracy of SMA-derived fractional cover
estimates depends highly on endmember purity, among other things, a single-endmember
model may fail to account for natural variability in the reflectance of endmembers.
Multiple Endmember Spectral Mixture Analysis (MESMA) addresses this issue by
allowing both the type and number of endmember models to vary on a per pixel basis
(Roberts et al. 1998; Asner & Lobell 2000). MESMA has been widely utilized for
deriving fractional cover estimates in several ecosystems using medium spatial resolution
(~20–30 m) multispectral (e.g. Landsat/ASTER) and hyperspectral (AVIRIS/ EO-1
Hyperion) data-sets (Roberts et al. 1998; Asner & Lobell 2000; Elmore et al. 2000;
Powell et al. 2007; Myint & Okin 2009). Although MESMA of hyperspectral images has
provided reliable estimates of fractional cover, it has produced modest results with
multispectral imagery due to non-ideal bandwidth and spatial resolution (Okin et al.
2001; Asner & Heidebrecht 2002; Gill & Phinn 2008). Besides spectral resolution, spatial
resolution of imagery also impacts the accuracy of MESMA-derived fractional estimates.
This is especially important in the semi-arid savanna ecosystems marked by both high
spatial heterogeneity and functional diversity in vegetation properties (Asner et al. 2011).
Such heterogeneity can significantly limit the accuracy of bio-physical estimates derived
from medium resolution multispectral imagery (e.g. Landsat) and may require data at a
higher spatial resolution. More recently, the availability of high spatial resolution (i.e.
<2.5 m) multispectral data-sets (e.g. GeoEye, Quickbird) presents new opportunities for
testing their suitability for fractional land cover mapping in such heterogeneous areas.
However, due to limited spectral dimensionality of high spatial resolution data-sets very
few studies have attempted such analysis (Hamada et al. 2011).
Since savanna ecosystems are considered patch dynamic systems (Wiegand et al.
2006; Meyer et al. 2009), mapping vegetation properties in savannas may benefit from
object-based image analysis (OBIA) (Laliberte et al. 2004; Blaschke 2010; Johansen
et al. 2010). Based on OBIA, imagery is first segmented into objects representing
 862 N.B. Mishra and K.A. Crews

relatively homogeneous group of pixels followed by sample selection and classification

of segments into classes of interest (Benz et al. 2004; Blaschke 2010). Important OBIA
advantages include (i) treatment of landscape to be consisting of relatively homoge-
neous patches (or ‘objects’) is ecologically more suitable than arbitrarily imposed pix-
els, (ii) landscape patches often depict scale dependency and based on OBIA,
landscape objects can be generated at multiple segmentation scales that can provide
added insight into ecological processes (Turner 2005) and (iii) OBIA approach offers
the opportunity to add contextual, geometrical and texture-related features (Kim et al.
2011). However, OBIA has not yet been extensively tested in natural landscapes where
continuous variation in vegetation features makes it difficult to determine boundaries
between ‘objects’ (Yu et al. 2008; Blaschke 2010). Further, in heterogeneous natural
systems, landscape patches can vary in size, shape and structure that may require char-
acterization at multiple spatial scales (Moustakas et al. 2009). Within the OBIA
approach, ‘segmentation scale’ parameter used in segmentation controls the maximum
Downloaded by [University of Boras] at 06:38 05 October 2014

allowable object-level spectral heterogeneity and object size (Trimble 2011a). Segmen-
tation scale is not prescriptive of the size (scale) of the resulting object, nor directly
related to the resolution (scale) of the input data. It can more suitably be described as
‘object complexity limit’, but the term ‘segmentation scale’ is well established and so it
is retained. Segmentation scale also affects the quality of segmented objects, which in
turn, affects the classification accuracy (Yu et al. 2006; Addink et al. 2007). By manip-
ulating the segmentation scale, nested objects of different size can be created that could
be further hierarchically classified representing the output at multiple scales. While such
hierarchical classification approach has been found suitable for characterizing spatially
heterogeneous systems such as wetlands (Dronova et al. 2012) and arid rangelands
(Laliberte et al. 2004), it’s suitability for characterizing fractional cover in dry savanna
systems merits further investigation. The objective of this study is to determine a better
image analysis approach for estimating fPV, fNPV and fBS in semi-arid savannas by com-
paring the suitability and limitations of MESMA and hierarchical OBIA approach
against in situ-obtained fractional cover estimates. The impact of vegetation structural
and function heterogeneity on the accuracy of fractional cover derived using these two
image analysis methods is also examined.

2. Study area
This study was conducted in the semi-arid central Kalahari region of Botswana in
southern Africa. The study area is approximately 17 km2 and falls completely within a
protected area (i.e. Central Kalahari Game Reserve) (Figure 1(a)). The rainfall in the
area is seasonal (between December and March) and the long-term annual average of
precipitation amounts to 350 mm usually with high spatio-temporal variability
(Makhabu et al. 2002; Scholes et al. 2002). Geologically, the area is dominated by
Kalahari sand with sporadic outcrops of calcrete (Moore & Attwell 1999). Vegetation
physiognomic properties in the study area are characterized by structurally
heterogeneous mixture of woody and herbaceous species that exhibit temporally distinct
phenological patterns (Figure 1(c)). Notable broad-leafed species in the area include
Lonchocarpus nelsii, Terminalia sericea, Bauhinia petersiana, Combretum hereroense
and Croton gratissimu. Important fine-leafed species are Acacia erioloba, Acacia
luederitzii, Ziziphus mucronata, Acacia mellifera and Acacia erubescens (Moore &
Attwell 1999). Vegetation boundaries based on plant species are often unclear since
differences among vegetation communities are related to changes in species dominance
 Geocarto International 863
Downloaded by [University of Boras] at 06:38 05 October 2014

Figure 1. (a) Location of study area within the central Kalahari of Botswana in Southern Africa,
(b) study area as seen from GeoEye-1 imagery (RGB:421), (c) enlarged parts of the study area
depicting different vegetation morphology types: (i) woodland, (ii) open shrubland, (iii) very open
shrubland, (iv) grassland and (v) pan.

rather than occurrence of different species (Moore & Attwell 1999; Makhabu et al.
2002). Pan areas are geomorphologically different from other parts of the study area
with flat topography and a high clay content soil. Pans are ecologically also important
as they attract large herbivore and associated predators (Parris & Child 1973).

3. Data and methods

3.1. Field data collection and development of vegetation morphology classes
In situ data on fractional cover and vegetation morphological characteristics were
acquired during the field campaign conducted in May 2012. Fractional cover of fPV,
fNPV and fBS were quantitatively estimated using a 30 m transect following the line-
intercept approach modified from Herrick et al. (2005). Each transect was divided into
segments of 50 cm (total 60 segments within one transect) for which fractional cover of
fPV, fNPV and fBS was visually estimated and recorded. The fraction of each cover type
 864 N.B. Mishra and K.A. Crews

was averaged from these 60 segments to get the fPV, fNPV and fBS for every transect.
Due to serious accessibility and safety issues (e.g. danger of predator attack), transects
could not be established away from existing tracks. Field data was collected in 18 total
transects spatially distributed across different vegetation morphology types. Geographic
coordinates were recorded at the start and end points of each transect using global posi-
tioning system equipment. Additional information included visual estimation of domi-
nant vegetation functional type (e.g. woody vs. herbaceous), minimum, maximum and
average vegetation height, dominant tree/shrub species and pictures acquired with a
digital camera. Based on the analysis of field data, five vegetation morphology classes
were developed mainly considering vegetation physiognomy, vertical and horizontal
agreement, leaf type and phenology. These classes were (i) Mixed deciduous woodland
with shrubs and herbaceous layer (Woodland), (ii) Mixed (70–40%) medium high
shrubland with open short herbaceous layer (Open shrubland), (iii) Mixed (40–10%)
medium high shrubland with open short herbaceous layer (Very open shrubland), and
Downloaded by [University of Boras] at 06:38 05 October 2014

(iv) Medium tall grassland with medium high shrubs (Grassland) and (v) Pans and bare
areas (Pans) (Figure 1(c)).
To avoid the influence of seasonality, field data on fractional cover was collected in
the same month as image acquisition (i.e. May). In spite of the two year temporal dif-
ference between image acquisition and in situ data collection, this study could directly
compare them because of the following evidence: (i) the study area lies completely
inside a protected area and does not have any anthropogenic influence, and (ii) due to
low animal density the grazing impact from herbivores is very minimal. Although, spa-
tio-temporal pattern of rainfall in the central Kalahari is highly variable, in this study
there are no means to quantify this because the nearest meteorological station (i.e.
Maun) is about 150 km north of the study area.

3.2. Remotely sensed imagery and pre-processing

This study utilized a multispectral GeoEye-1 imagery consisting 2-m blue, green, red
and near-infrared bands acquired on 16 May 2010 at 10:46 local time (08:46 UTC). In
the central Kalahari, May represents the beginning of the dry season when trees and
shrubs are still green whereas herbaceous vegetation contained no observable green bio-
mass. Surface reflectance was derived by radiometrically calibrating the imagery using
ATCOR algorithm. ATCOR is an absolute atmospheric correction method that applies
an atmospheric look-up table based on a large database containing the result of radia-
tive transfer calculation from the MODTRAN-4 radiative transfer code (Richter &
Schläpfer 2008). The optical depth of the atmospheric aerosols was calculated by com-
paring modelled at-sensor radiance with measured radiance in the red band of areas
with dark dense vegetation. This correction was then applied on each pixel to derive
surface reflectance. As an additional feature Normalized Difference Vegetation Index
(NDVI) was derived from the calibrated imagery.

3.3. Multiple Endmember Spectral Mixture Analysis

The objective of MESMA analysis was to derive sub-pixel estimates of fPV, fNPV and
fBS from the GeoEye imagery. NDVI calculated from the GeoEye imagery was found
to be weakly correlated with the red band (R2 = 0.41) and very weakly correlated with
the NIR band (R2 = 0.12). Thus, the NDVI was considered as a non-linear combination
of red and NIR bands and was stacked with the existing GeoEye bands before
 Geocarto International 865

endmember selection and MESMA analysis. Endmembers correspond to the spectra of

spectrally pure materials and the goal of endmember selection process is to isolate such
pure pixels by utilizing qualitative and/or quantitative purity measures. In this study,
endmember candidates were first selected based on: (i) pixel purity index values, and
(ii) visualization of the multidimensional feature space plots and spectral indices values
(Bateson & Curtiss 1996; Plaza et al. 2004). These candidate spectra were analysed
quantitatively using three fit matrices: Endmember Average RMSE to select endmem-
bers that produced the lowest RMSE within a class (Dennison & Roberts 2003), Mini-
mum Average Spectral Angle that isolates endmembers with the lowest average
spectral angle (Dennison et al. 2004) and Count-Based endmember selection (Roberts
et al. 2003). Finally, for PV, NPV and soil six, eight and six endmembers were
selected, respectively (Figure 2).
MESMA was performed using the VIPER tools (www.vipertools.org) programme
which is a free add-on to ENVI image processing software. Non-shade endmember
Downloaded by [University of Boras] at 06:38 05 October 2014

fraction was calculated using singular value decomposition and shade was calculated
as 1 minus the sum of all non-shade endmember fractions. Three model schemes
(two-, three- and four-endmembers) were tested for each pixel (Table 1). MESMA
was partially constrained with minimum and maximum allowable non-shade frac-
tions and the RMSES threshold set to −0.05, 1.05 and 0.025, respectively. Here,
RMSES refers to the root-mean-square error or the model residual used to access
the model fit. For choosing the best mixing model for each pixel, first, the model
producing the lowest RMSES was selected as the best model for each pixel at each
model complexity level. In the second step, output composites of two-, three- and
four-endmember models were compared. Since a three-endmember model will
always produce a lower RMSES than a two-endmember model (same for four- vs.
three-endmember model results) the following criteria was adopted for their compari-
son: (i) if a three-endmember model had a lower RMSES than a two-endmember
model and the three-endmember model exceeded a predefined threshold of decreased
RMSES (0.007, determined empirically in this study similar to Powel et al. 2007),
then the three-endmember model was considered superior; otherwise, the two-end-
member model was selected as superior and (ii) If a four-endmember model had a
lower RMSES than the best three-endmember model and the four-endmember model
exceeded the same threshold, then the four-endmember model was selected;
otherwise the three-endmember model was considered superior. MESMA-obtained
fractional cover estimates were shade normalized where the shade fraction from the
results was taken out by dividing the fraction of each non-shade endmember by the
sum of non-shade endmembers.

Figure 2. Spectral profile of endmembers used for MESMA analysis in this study.
 866 N.B. Mishra and K.A. Crews

Table 1. Different types and varying complexity of MESMA models used in this study. The
number besides each model types represents the number of models tested at each model complex-
ity level.

Two-endmember models Three-endmember models Four-endmember models

PV + Shade (6) PV + NPV + Shade (48) PV + NPV + Soil + Shade (288)
NPV + Shade (8) PV + Soil + Shade (36)
Soil + Shade (6) Soil+ NPV + Shade (48)
Total models: 440

3.4. Object-based image analysis

The objective of object-based analysis was to derive estimates of fPV, fNPV and fBS
Downloaded by [University of Boras] at 06:38 05 October 2014

from the GeoEye-1 imagery. To implement object-based classification, first, the ras-
ter stack (4 GeoEye bands and NDVI) was segmented using the multiresolution
segmentation algorithm in eCognition Developer 8 software (Trimble 2011b). Com-
pared to other segmentation approaches, the multiresolution segmentation allows
construction of objects of different sizes and enhances the response of object gener-
ation to landscape patch structure (Baatz & Schäpe 2000; Arbiol et al. 2006). Due
to spatial and structural heterogeneity in the study area, meaningful spectrally
homogeneous objects can occur at different spatial scales (Arbiol et al. 2006).
Additionally, it was expected that optimal patch size would vary depending on veg-
etation physiognomy. Hence, objects were generated at five segmentation levels at
scale values 1, 2, 4, 6 and 8 (Figure 3). The other required segmentation parame-
ters, shape and compactness were kept at a constant value of 0.2 and 0.4, respec-
tively (Table 2). These values were selected as optimal, based on visual
comparison of the results of several iterative segmentation runs. The input imagery
bands were given equal weight while the NDVI was given higher weight compared
to other inputs. For classification at each segmentation scale, training objects were
selected by visual interpretation.
For the selection of optimal classification features, initially 42 spectral, shape,
texture and geometrical features were selected to include a sufficiently wide range.
First, Spearman’s rank correlation analysis was utilized to eliminate features with
correlation coefficients above 0.9. Seventeen features were found to have correla-
tions below this threshold value. The selection of the best features from these was
based on Jeffrey’s-Matusita distance (JM distance) calculated using the SEATH tool
(Nussbaum et al. 2006; Marpu et al. 2008). This study calculated the largest aver-
age JM distance (the mean of all two class combinations) for every possible 12–17
feature combination. Combination with lesser number of features always had a
lower JM distance compared to combinations with more features and selecting the
lowest mean JM distance for the 11 candidates was deemed unsuitable. Hence, the
feature combination that resulted in the largest JM distance for the least separable
pair of classes was selected which is also an approach followed in previous studies
(Swain & Davis 1978; Laliberte et al. 2012). This approach resulted in the final
selection of 13–17 features depending on the segmentation scale factor (Table 3).
Final classification at each of the segmentation scales was performed using these
optimal features and object samples for each class using the nearest neighbour
classification technique within eCognition.
 Geocarto International 867
Downloaded by [University of Boras] at 06:38 05 October 2014

Figure 3. Illustration of how optimal segmentation scale following OBIA approach varies for
different vegetation morphology types in the semi-arid central Kalahari.

Table 2. Segmentation parameters and associated object statistics for the five segmentation
scales considered in the study.

Segmentation Colour/ Smoothness/ Minimum/mean/maximum Total # of

Level scale shape compactness number of pixels in objects objects
1 1 0.8/0.2 0.6/0.4 1/1.4/8 4,119,396
2 2 0.8/0.2 0.6/0.4 1/2.1/25 3,698,076
3 4 0.8/0.2 0.6/0.4 1/3.37/164 1,567,399
4 6 0.8/0.2 0.6/0.4 1/7.53/400 631,129
5 8 0.8/0.2 0.6/0.4 2/16.7/1105 316,343

3.5. Fractional cover validation/spatial association with vegetation morphology

classes
In this study, in situ-derived fractional cover estimates were first used to evaluate the
accuracy of estimates derived from both MESMA and OBIA approaches. For this pur-
pose, the fractional estimates of each 30 m transect were compared with the shade-nor-
malized mean fPV, fNPV and fBS of spatially coincident 225 pixels (15 × 15 pixel
window) equivalent to the transect length. Root-mean-square error (RMSEC) was used
as the error matrix and was calculated based on considering in situ-derived estimates as
 868 N.B. Mishra and K.A. Crews

Table 3. Overview of the selected features used in image classification based on the OBIA
approach.

Features Description
Mean layer value The mean value represents the mean brightness of an image object within a
single band; used feature: Mean NIR, Mean NDVI
Ratios The amount that a given band contributes to the total brightness; used
feature: Ratio NDVI, Ratio Blue, Ratio Red
Standard The standard deviation of all pixels which form an image object within a
deviation band; used feature: Standard Deviation Green, Standard Deviation NDVI,
Standard Deviation Blue, Standard Deviation NIR, Standard Deviation
GLCM (all bands)
Maximum Minimum mean value of an object subtracted from its maximum value. The
difference means of all bands belonging to an object are compared with each other and
the result is divided by the brightness
Texture after GLCM (Grey Level Co-occurrence Matrix) calculated after Haralick et al.
Downloaded by [University of Boras] at 06:38 05 October 2014

Haralick (1973): Describes how different combinations of pixel values occur within an
object; used features (using the mean of all layers): GLCM Mean, GLCM
Contrast, GLCM Dissimilarity, GLCM Standard Deviation, GLDV (Grey
Level Difference Vector): The sum of the diagonals of the grey level co-
occurrence matrix; used features (using the mean of all layers): GLDV
entropy

observed and MESMA- and OBIA-derived estimates as predicted estimates. The sub-
script ‘c’ (as opposed to subscript ‘s’ used above) refers to the root-mean-square error
calculated in the comparison of cover estimates. However, due to the limited number
of field transects (n = 18) there was need for additional evaluation. Both qualitative and
quantitative evaluation of results suggested that MESMA produced reliable fractional
estimates than OBIA. Hence, as a second evaluation step, MESMA-derived estimates
were used to evaluate the OBIA-derived fractional estimates obtained at the five differ-
ent segmentation scales. For this purpose, a total of 214 grid cells each covering 30 ×
30 m area (225 pixels) were created and their shade-normalized mean fPV, fNPV and fBS
derived from MESMA and OBIA were compared. Results were first evaluated by con-
sidering all 214 samples together to access overall agreement and then by grouping
samples based on their vegetation morphology for examining their spatial association.

4. Results
4.1. Field vs. MESMA- and OBIA-derived estimates
Fractional estimates derived in situ and those estimated for corresponding areas from
MESMA and OBIA of imagery are plotted in Figure 4. Comparison of in situ transect-
derived fractional estimates with those obtained from MESMA and the hierarchical
OBIA approach depicted that MESMA produced a lower error for all ground cover
component types compared to OBIA-derived estimates (i.e. overall RMSEC: 6.2%)
(Table 4). Furthermore, based on the OBIA approach, the lowest error in fractional
cover estimates were obtained at the finest segmentation scale (i.e. scale = 1, overall
RMSEC 8.6%) and with increasing segmentation scale, error increased consistently for
all ground cover component types reaching the highest value at the segmentation scale
value of 8 (i.e. overall RMSEC: 22.9%) (Table 4). MESMA-derived fractional cover
estimates also showed cover specific differences in estimation accuracy as fPV could be
 Geocarto International 869
Downloaded by [University of Boras] at 06:38 05 October 2014

Figure 4. Comparison of field estimated fractional cover of fPV, fNPV and fBS against those esti-
mated from MESMA and OBIA. The OBIA fractional estimates presented here were derived at
segmentation scale value.

Table 4. Evaluation results of MESMA and hierarchical OBIA-derived fractional estimates

against field derived.

RMSEC: All samples (n = 18)

Technique/segmentation scale
fPV (%) fNPV (%) fBS (%) Overall (%)
MESMA 5.1 7.1 7.4 6.5
OBIA (segmentation scale = 1) 7.6 7.8 10.6 8.6
OBIA (segmentation scale = 2) 10.4 11.1 13.8 11.7
OBIA (segmentation scale = 4) 13.7 16.6 17.2 15.8
OBIA (segmentation scale = 6) 19 18.1 18.8 18.6
OBIA (segmentation scale = 8) 21.8 22.3 24.6 22.9

estimated with lower error (i.e. RMSEC: 5.1%) compared to fNPV and fBS (RMSEC: 7.1
and 7.4%, respectively). Based on OBIA, similar cover specific differences in fractional
cover estimation accuracy were observed at the finest segmentation scale ( fPV RMSEC:
7.6% vs. fBS RMSEC: 10.6%). However, with the increasing segmentation scale in
OBIA, these cover specific differences became indistinguishable as the estimation error
increased consistently reaching a maximum value at the segmentation scale value 8
(Table 4).

4.2. MESMA- vs. OBIA-derived fractional estimates: agreement across vegetation

morphology types
The results of fractional cover mapping based on the MESMA and hierarchical OBIA
approach for the five different regions of the study area with different vegetation
morphology types are shown in Figure 5. Evaluation of OBIA-derived fractional cover
 870 N.B. Mishra and K.A. Crews
Downloaded by [University of Boras] at 06:38 05 October 2014

Figure 5. MESMA vs. hierarchical OBIA classification results for areas under five different
vegetation morphology types in the central Kalahari.

estimates considering MESMA results as observed estimates depicted several

interesting results that are reported in Table 5. Based on OBIA, fractional estimates
derived at the finest segmentation scale (i.e. value 1) produced the lowest overall error
(i.e. RMSEC 7.7%) as well as the lowest cover specific error in fractional cover
(Table 5(f )). With increasing segmentation scale value, error in fractional cover estima-
tion increased consistently. Furthermore, analysis of the results by grouping samples
based on the vegetation morphology type showed that observed sensitivity of hierarchi-
cal OBIA results were also vegetation morphology specific. For all vegetation morphol-
ogy classes, the lowest error in fractional cover estimates was achieved at the finest
segmentation scale (segmentation scale = 1). However, even at the finest segmentation
scale, estimation accuracy varied depending on morphology type. For example, frac-
tional estimates in areas representing vegetation morphology class 1 and class 2 were
mapped with low error (overall RMSEC 4.2 and 5.1%, respectively) (Table 5(a) and
(b)) compared to areas under morphology class 4 or 5 that showed higher error (overall
RMSEC 10.4 and 11.6%, respectively) (Table 5(d) and (e)). Furthermore, with increas-
ing segmentation scale, different vegetation morphology types showed different
responses in terms of fractional cover estimation accuracy. Areas under vegetation mor-
phology class 1 were less sensitive to changing segmentation scale and fractional cover
estimation error was comparatively lower for these areas even at coarser segmentation
scales (i.e. overall RMSEC at segmentation scale 8: 10.2%). On the contrary, areas
under vegetation morphology class 4 and 5 were more sensitive to changing segmenta-
tion scale as the increase in the magnitude of estimation error was much higher at
 Geocarto International 871

Table 5. Evaluation of hierarchical OBIA approach-derived fractional estimates considering

MESMA-derived estimates as observed estimates. 5(a)–(e) represent results based on vegetation
morphology type of samples where (a) Woodland, (b) Open shrubland, (c) Very open shrubland,
(d) Grassland, (e) Pans and (f ) represents the overall results considering all samples.

fPV fNPV fBS Overall fPV fNPV fBS Overall

Scale (%) (%) (%) (%) Scale (%) (%) (%) (%)
(a) RMSEC: class 1 woodland (n = 28) (d) RMSEC: class 4 grassland (n = 43)
1 4.3 3.6 4.9 4.2 1 9.1 11.8 10.3 10.4
2 4.9 4.3 6.2 5.1 2 14.8 13.2 13.8 13.9
4 5.5 5 7.7 6 4 17.5 14.1 15.5 15.7
6 6.8 8.1 9.1 8 6 20.1 18 21.5 19.8
8 7.9 10.4 12.3 10.2 8 24.6 20.3 24.4 23.1
(b) RMSEC: Class 2 Open shrubland (e) RMSEC: Class 5 Pans (n = 55)
(n = 40)
1 5.1 4.8 5.5 5.1 1 9.9 13.8 11.1 11.6
Downloaded by [University of Boras] at 06:38 05 October 2014

2 7.4 6.8 7.1 7.1 2 11 14.3 13.4 12.9

4 8.8 8.1 11.9 9.6 4 12.2 17.8 14.9 14.9
6 10.3 10.1 14.4 11.6 6 14.3 21.4 16.5 17.4
8 14.3 14.4 17 15.2 8 17.7 24.8 21.6 21.3
(c) RMSEC: Class 3 Very open (f) RMSEC: All samples (n = 214)
shrubland (n = 48)
1 5.9 7.1 9.1 7.3 1 6.8 8.3 8.1 7.7
2 6.6 8.8 13.5 9.6 2 8.9 9.4 10.8 9.7
4 9.2 11.3 16.3 12.2 4 10.6 11.1 13.2 11.6
6 16.1 13.4 17.7 15.7 6 14.2 14.2 15.8 14.7
8 21 18.4 23.2 20.8 8 17.1 17.6 19.7 18.1

higher segmentation scale value (e.g. for class 4 and class 5, the overall RMSEC at seg-
mentation scale 8: 32.1 and 21.3%, respectively).

5. Discussion
5.1. Field vs. MESMA- and OBIA-derived estimates
Results of this study suggest that when compared to in situ-derived estimates, MESMA
produced lower error than OBIA for fractional cover estimation in the semi-arid central
Kalahari savanna. These results could be explained in light of how the image analysis
approaches compared in this study (i.e. MESMA vs. OBIA) represent and model the
inherent structural and functional heterogeneity in semi-arid systems as captured by the
GeoEye-1 imagery. Due to its high spatial resolution, GeoEye-1 imagery is able to dis-
tinguish individual tree/shrub canopies and also captures intra- and inter-canopy details,
e.g. sunlit vs. shaded areas, green vs. senescent foliage (Wulder et al. 2009). Given the
high spatial heterogeneity in the study area, mixtures of different cover types (espe-
cially of NPV and soil) can exist even within a GeoEye-1 pixel. The MESMA
approach attempts to account for this fine-scale heterogeneity and provides sub-pixel
abundance estimates by modelling a pixel’s spectral response as a linear combination
of considered endmember types. In heterogeneous savanna systems, the potential of
non-linear mixing is higher at higher spatial resolution compared to lower spatial reso-
lution. The endmembers used for MESMA analysis were thoroughly tested using both
qualitative and quantitative purity measures. Using these pure and representative
 872 N.B. Mishra and K.A. Crews

endmember spectra that were allowed to vary in number and type on a per-pixel basis,
all pixels of the GeoEye-1 imagery could be modelled within 2.5% RMSEs and pro-
ducing accurate fractional cover estimates. Qualitative assessments of MESMA-derived
fractional estimate against photo acquired at transect location showed good correspon-
dence. Due to its high spatial resolution the input GeoEye imagery could be used for
qualitative assessment which also depicted good agreement across all vegetation mor-
phology types. Unlike MESMA, the OBIA approach considers the landscape to be con-
sisting of relatively homogeneous objects. At the finest segmentation scale used in this
study (i.e. scale value 1), an average object consists of nearly the size of a single Geo-
Eye-1 pixel and with increasing segmentation scale the object size (and the number of
pixels within an object) increases (as depicted in Table 2). In this study, OBIA results
at the finest segmentation scale produced a lower estimation error because it better rep-
resented the landscape heterogeneity compared to higher segmentation scale-derived
estimates. Increasing segmentation scale value under OBIA resulted in a larger object
Downloaded by [University of Boras] at 06:38 05 October 2014

size (Figure 3) that is an oversimplified representation of landscape heterogeneity. Clas-

sification of these larger objects produced a much higher estimation error compared to
both OBIA results at finer segmentation scale as well as MESMA-derived estimates.

5.2. MESMA- vs. OBIA-derived fractional estimates: Agreement across vegetation

morphology types
The results of this study found that variations in savanna vegetation morphological
properties impact fractional cover estimation accuracy. The varying sensitivity in the
fractional cover estimation accuracy of different vegetation morphology types to OBIA
segmentation scale could be explained by the varying patch sizes of savanna vegetation
in the central Kalahari. Patch is the basic element of landscape structure and can be
defined as an ecologically homogeneous unit. Patches are homogeneous at a particular
scale and can become heterogeneous at increasingly coarser scales (Kotliar & Wiens
1990). Patch size in savanna systems has been found to be closely associated with veg-
etation physiognomic properties (Meyer et al. 2009). Area dominated and co-dominated
by woody life forms (i.e. trees and/or shrubs) (e.g. woodlands and open shrublands)
have larger patch sizes compared to areas that are dominated by herbaceous compo-
nents (e.g. grasslands, very open shrublands) (Figure 1(c)). Thus, due to their larger
patch size, areas under woodlands and open shrublands (i.e. morphology class 1 and 2,
respectively) produce homogeneous objects in terms of ground cover components at
both small and large segmentation scale values and yield more accurate fractional esti-
mates. On the contrary, grassland and very open shrubland areas (i.e. morphology class
4 and 3) have a much smaller patch size. In these areas, homogeneous objects are cre-
ated only when image segmentation is performed at a fine segmentation scale resulting
in small objects. At high segmentation scale values, image objects for these areas have
more than one ground cover component (e.g. mixture of fNPV and fBS). The classifica-
tion result of these large and comparatively mixed objects results in a high error in
fractional cover when compared to either in situ- or MESMA-derived estimates. These
results underscore that besides the mapping method, the understanding of spatial and
structural heterogeneity is also important in semi-arid savanna systems. To a certain
extent, vegetation structure/land cover characteristics in the study area can be treated as
representative of other naturally occurring semi-arid savanna systems found in southern
Africa (Dougill et al. 1999) and tropical savannas in other parts of the world (e.g. Aus-
tralia, South America). Thus, these results have several implications for future studies
 Geocarto International 873

focusing on characterization and monitoring of semi-arid savanna systems using com-

bined in situ data and high spatial resolution imagery. Future studies utilizing high reso-
lution imagery for quantifying ecosystem-related functional properties (e.g. fractional
cover) in savannas should select the image analysis approach considering the structural/
functional heterogeneity of the study area.
There are some sources of uncertainty in the data and method utilized in this study.
The temporal difference between field data collection (i.e. May 2012) and image acqui-
sition (i.e. May 2010) is likely to reduces the accuracy of imagery-derived estimates to
some extent as vegetation dynamics in central Kalahari is affected by spatio-temporal
variability in rainfall (Scholes et al. 2002; Scanlon et al. 2005). Additionally, the visual
separation of fPV and fNPV in the field was difficult and subjective to some extent. Thus,
even with best efforts and careful interpretation, in situ-derived fractional estimates con-
tain some unquantifiable observer error. Furthermore, MESMA assumes linearity in the
mixing of endmember spectra. However, in semi-arid systems complex vegetation
Downloaded by [University of Boras] at 06:38 05 October 2014

structure leads to non-linear mixing that may limit the accuracy of MESMA to a certain
extent (Okin & Roberts 2004).

6. Conclusions
This study compared absolute estimates of fPV, fNPV, fBS derived in situ and from Geo-
Eye-1 imagery based on MESMA and OBIA approach to examine the comparative
suitability of these methods. Comparison of analysis methods of remote sensing imag-
ery is informative. Given the same input imagery comparison of results from different
methods provides information on the inherent limitations among methods, regardless of
in situ data. We also analysed the association of results obtained from these methods
with vegetation morphology types in the study area. Our results show that in the xeric
central Kalahari, MESMA of GeoEye-1 imagery produced more accurate fractional
cover estimates compared to hierarchical OBIA approach. Notably, OBIA results at the
finest segmentation scale were close to field estimtates as well as MESMA-derived esti-
mates. The sub-pixel analysis approach of MESMA was able to represent the spatial
heterogeneity in semi-arid central Kalahari better than the OBIA approach. However,
an important limitation of MESMA is the high computational time and computing
resources required due to the large volume of high spatial resolution imagery making it
less suitable for landscape or regional scale applications. Nonetheless, MESMA results
for strategically selected representative sample areas could be used to validate results
derived using coarse spatial resolution data that provide large spatio-temporal coverage
but might be less accurate. While field data are critical for evaluating remote sensing
methods, it also has inherent errors and its acquisition, especially in remote areas can
be highly limited due to logistical and safety issues. Due to the limited number of field
transects, this study used MESMA-derived fractions to validate OBIA results. The fine
spatial detail provided by high spatial resolution imagery and its automated analysis
using MESMA has the potential to substitute the requirement of field data to some
extent and thus serve as a pragmatic approach for accurately characterizing savanna
ecosystems.

Acknowledgements
This research was supported by National Science Foundation: Doctoral Dissertation improvement
grant #1203580 and the Veselka field research grant from University of Texas at Austin. GeoEye
 874 N.B. Mishra and K.A. Crews

imagery was provided as imagery grant from the GeoEye Foundation. We would like to thank
Thoralf Meyer and Glyn Maude for field work planning and logistical support, field assistants for
help and guidance during field campaigns, Mario Cardozo and Gargi Chaudhuri for discussions
that helped improve this work. We would like to thank the two reviewers whose comments help
improve the overall quality of the manuscript.

References
Adams JB, Sabol DE, Kapos V, Almeida R, Roberts DA, Smith MO, Gillespie AR. 1995. Classi-
fication of multispectral images based on fractions of endmembers: application to land-cover
change in the brazilian amazon. Remote Sens Environ. 52:137–154. Available from: <Go to
ISI>://WOS:A1995RE08800005
Addink EA, De Jong SM, Pebesma EJ. 2007. The importance of scale in object-based mapping
of vegetation parameters with hyperspectral imagery. Photogrammetric Eng Remote Sens.
73:905–912.
Arbiol R, Zhang Y, Palá V. 2006. Advanced classification techniques: a reviewed. ISPRS Com-
Downloaded by [University of Boras] at 06:38 05 October 2014

mission VII Mid-term Symposium From Pixel to Processes, Enschede, NL.

Asner GP, Heidebrecht KB. 2002. Spectral unmixing of vegetation, soil and dry carbon cover in
arid regions: comparing multispectral and hyperspectral observations. Int J Remote Sens.
23:3939–3958. Available from: <Go to ISI>://000178104900004
Asner GP, Levick SR, Smith PJ. 2011. Remote sensing of fractional cover and biochemistry in
savannas. In: Hill MJ, Hanan NP, editors. Ecosystem function in savannas: measurement and
modeling at landscape to global scales. Boca Raton (FL): CRC Press; p. 195–218.
Asner GP, Lobell DB. 2000. A biogeophysical approach for automated swir unmixing of soils
and vegetation. Remote Sens Environ. 74:99–112. Available from: http://www.sciencedirect.
com/science/article/B6V6V-4105CWR-8/2/7bbafe8846994c7433c175e9d2d0b7a2
Baatz M, Schäpe M. 2000. Multiresolution segmentation – an optimization approach for
high-quality multi-scale image segmentation. In: Strobl J, Blaschke T, Griesebner G, editors.
Angewandte geographische informations-verarbeitung. Karlsruhe: Wichmann Verlag; p.
12–23.
Bateson A, Curtiss B. 1996. A method for manual endmember selection and spectral unmixing.
Remote Sens Environ. 55:229–243.
Benz UC, Hofmann P, Willhauck G, Lingenfelder I, Heynen M. 2004. Multi-resolution, object-
oriented fuzzy analysis of remote sensing data for gis-ready information. ISPRS J Photo-
gramm Remote Sens. 58:239–258. Available from: <Go to ISI>://000188924400008
Blaschke T. 2010. Object based image analysis for remote sensing. ISPRS J Photogramm Remote
Sens. 65:2–16. Available from: <Go to ISI>://WOS:000274448200002
Blaum N, Rossmanith E, Jeltsch F. 2007. Land use affects rodent communities in Kalahari savan-
nah rangelands. Afr J Ecol. 45:189–195. Available from: <Go to ISI>://000246576600011
Dennison PE, Halligan KQ, Roberts DA. 2004. A comparison of error metrics and constraints for
multiple endmember spectral mixture analysis and spectral angle mapper. Remote Sens Envi-
ron. 93:359–367.
Dennison PE, Roberts DA. 2003. Endmember selection for multiple endmember spectral mixture
analysis using endmember average RMSE. Remote Sens Environ. 87:123–135. Available
from: <Go to ISI>://000186447400001
Dougill AJ, Thomas DSG, Heathwaite AL. 1999. Environmental change in the Kalahari: inte-
grated land degradation studies for nonequilibrium dryland environments. Ann Assoc Am
Geographers. 89:420–442. Available from: http://www.jstor.org/stable/2564491
Dronova I, Gong P, Clinton NE, Wang L, Fu W, Qi S, Liu Y. 2012. Landscape analysis of wet-
land plant functional types: the effects of image segmentation scale, vegetation classes and
classification methods. Remote Sens Environ. 127:357–369. Available from: <Go to ISI>://
BCI:BCI201300129817
Elmore AJ, Mustard JF, Manning SJ, Lobell DB. 2000. Quantifying vegetation change in semiarid
environments. Remote Sens Environ. 73:87–102. Available from: http://www.sciencedirect.
com/science/article/B6V6V-40NMT6N-8/2/fe897af662c76385be1dd872eb38b781
Gill T, Phinn S. 2008. Estimates of bare ground and vegetation cover from advanced spaceborne
thermal emission and reflection radiometer (aster) short-wave-infrared reflectance imagery. J
Appl Remote Sens. 2:023511–19. Available from: http://link.aip.org/link/?JRS/2/023511/1
 Geocarto International 875

Guerschman JP, Hill MJ, Renzullo LJ, Barrett DJ, Marks AS, Botha EJ. 2009. Estimating frac-
tional cover of photosynthetic vegetation, non-photosynthetic vegetation and bare soil in the
australian tropical savanna region upscaling the eo-1 hyperion and modis sensors. Remote
Sens Environ. 113:928–945. Available from: <Go to ISI>://000265716000006
Haralick, RM, Shanmuga, K, Dinstein, I. 1973. Textural features for image classification. IEEE T
Syst Man Cyb, SMC3 (6):610–621.
Hamada Y, Stow DA, Roberts DA. 2011. Estimating life-form cover fractions in california sage
scrub communities using multispectral remote sensing. Remote Sens Environ. 115:3056–
3068. Available from: <Go to ISI>://WOS:000298311300009
Herrick JE, Van Zee JW, Havstad KM, Burkett LM, Whitford WG. 2005. Monitoring manual for
grassland, shrubland and savanna ecosystems. Las Cruces (NM): USDA-ARS Jornada Exper-
imental Range.
Hill MJ, Roman MO, Schaaf CB. 2011. Biogeography and dynamis of global tropical and sub-
tropical savannas: a spatiotemporal view. In: Hill MJ, Hanan NP, editors. Ecosystem function
in savannas: measurement and modeling at landscape to global scales. Boca Raton (FL):
CRC Press; p. 3–38.
Johansen K, Arroyo LA, Armston J, Phinn S, Witte C. 2010. Mapping riparian condition indica-
Downloaded by [University of Boras] at 06:38 05 October 2014

tors in a sub-tropical savanna environment from discrete return lidar data using object-based
image analysis. Ecol Indic. 10:796–807. Available from: <Go to ISI>://
WOS:000276540000003
Kim M, Warner TA, Madden M, Atkinson DS. 2011. Multi-scale geobia with very high spatial
resolution digital aerial imagery: scale, texture and image objects. Int J Remote Sens.
32:2825–2850. Available from: <Go to ISI>://WOS:000290985000010
Kotliar NB, Wiens JA. 1990. Multiple scales of patchiness and patch structure: a hierarchical
framework for the study of heterogeneity. Oikos. 59:253–260. Available from: http://www.
jstor.org/stable/3545542
Lal R. 2004. Carbon sequestration in dryland ecosystems. Environ Manage. 33:528–544. Avail-
able from: http://dx.doi.org/10.1007/s00267-003-9110-9
Laliberte A, Browning D, Rango A. 2012. A comparison of three feature selection methods for
object-based classification of sub-decimeter resolution ultracam-l imagery. Int J Appl Earth
Obs Geoinf. 15:70–78.
Laliberte AS, Rango A, Havstad KM, Paris JF, Beck RF, Mcneely R, Gonzalez AL. 2004.
Object-oriented image analysis for mapping shrub encroachment from 1937 to 2003 in south-
ern New Mexico. Remote Sens Environ. 93:198–210. Available from: <Go to ISI>://
000224294400016
Ludwig JA, Bastin GN, Chewings VH, Eager RW, Liedloff AC. 2007. Leakiness: a new index
for monitoring the health of arid and semiarid landscapes using remotely sensed vegetation
cover and elevation data. Ecol Indic. 7:442–454. Available from: <Go to ISI>://
000244585600018
Makhabu SW, Marotsi B, Perkins J. 2002. Vegetation gradients around artificial water points in
the central Kalahari game reserve of botswana. Afr J Ecol. 40:103–109. Available from: <Go
to ISI>://000175478500001
Marpu PR, Nussbaum S, Niemeyer I, Gloaguen R. 2008. A procedure for automatic object-based
classification. In: Blaschke T, Lang S, Hay G, editors. Object-based image analysis spatial
concepts for knowledge-driven remote sensing applications series: lecture notes in geoinfor-
mation and cartography. Berlin: Springer; p. 169–184.
Meyer KM, Wiegand K, Ward D. 2009. Patch dynamics integrate mechanisms for savanna tree-
grass coexistence. Basic Appl Ecol. 10:491–499. Available from: <Go to ISI>://BIOSIS:
PREV200900638540
Mishra NB, Crews KA. 2014. Mapping vegetation morphology types in a dry savanna ecosys-
tem: integrating hierarchical object-based image analysis with random forest. Int J Remote
Sens. 35:1175–1198.
Mistry J. 2000. Savannas. Prog Phys Geogr. 24:601–608. Available from: <Go to ISI>://
000165678600008
Moore A, Attwell C. 1999. Geological controls on the distribution of woody vegetation in the
central Kalahari, botswana. S Afr J Geol. 102:350–362.
 876 N.B. Mishra and K.A. Crews

Moustakas A, Sakkos K, Wiegand K, Ward D, Meyer KM, Eisinger D. 2009. Are savannas
patch-dynamic systems? A landscape model Ecol Modell. 220:3576–3588. Available from:
<Go to ISI>://000272957300012
Myint SW, Okin GS. 2009. Modelling land-cover types using multiple endmember spectral mix-
ture analysis in a desert city. Int J Remote Sens. 30:2237–2257. Available from: <Go to
ISI>://000266917400005
Nussbaum S, Niemeyer I, Canty MJ. 2006. Seath – a new tool for automated feature extraction
in the context of object-oriented image analysis. Proceedings of AGIT – Obia; Salzburg: IS-
PRS.
Okin GS, Parsons AJ, Wainwright J, Herrick JE, Bestelmeyer BT, Peters DC, Fredrickson EL.
2009. Do changes in connectivity explain desertification? Bioscience. 59:237–244. Available
from: <Go to ISI>://000263863400008
Okin GS, Roberts DA. 2004. Remote sensing in arid regions: challenges and opportunities. Man-
ual Remote Sens. 4:111–146.
Okin GS, Roberts DA, Murray B, Okin WJ. 2001. Practical limits on hyperspectral vegetation
discrimination in arid and semiarid environments. Remote Sens Environ. 77:212–225. Avail-
able from: <Go to ISI>://000170461300008
Downloaded by [University of Boras] at 06:38 05 October 2014

Parris R, Child G. 1973. The importance of pans to wildlife in the Kalahari and the effect of
human settlements on these areas. J S Afr Wildlife Manage Assoc. 3:1–8.
Plaza A, Martinez P, Perez R, Plaza J. 2004. A quantitative and comparative analysis of end-
member extraction algorithms from hyperspectral data. IEEE Trans Geosci Remote Sens.
42:650–663.
Powell RL, Roberts DA, Dennison PE, Hess LL. 2007. Sub-pixel mapping of urban land cover
using multiple endmember spectral mixture analysis: Manaus, Brazil. Remote Sens Environ.
106:253–267. Available from: http://www.sciencedirect.com/science/article/B6V6V-4M7CDM
1-1/2/5c5dfce8a0276a2394385a98118569d3
Richter R, Schläpfer D. 2008. Atmospheric/topographic correction for satellite imagery. Atcor-2/3
user guide. Version 6.4.
Roberts DA, Dennison PE, Gardner ME, Hetzel Y, Ustin SL, Lee CT. 2003. Evaluation of the
potential of hyperion for fire danger assessment by comparison to the airborne visible/infrared
imaging spectrometer. IEEE Trans Geosci Remote Sens. 41:1297–1310.
Roberts DA, Gardner M, Church R, Ustin S, Scheer G, Green RO. 1998. Mapping chaparral in
the santa monica mountains using multiple endmember spectral mixture models. Remote Sens
Environ. 65:267–279. Available from: http://www.sciencedirect.com/science/article/B6V6V-41
05CWR-8/2/7bbafe8846994c7433c175e9d2d0b7a2
Roberts DA, Smith MO, Adams JB. 1993. Green vegetation, nonphotosynthetic vegetation, and
soils in AVIRIS data. Remote Sens Environ. 44:255–269. Available from: <Go to ISI>://
A1993LB11800010
Roy DP, Boschetti L, Giglio L. 2011. Remote sensing of global savanna fire occurence, extent
and properties. In: Hill MJ, Hanan NP, editors. Ecosystem function in savannas: measure-
ment and modeling at landscape to global scales. Boca Raton (FL): CRC Press; p. 220–239.
Scanlon TM, Caylor KK, Manfreda S, Levin SA, Rodriguez-Iturbe I. 2005. Dynamic response of
grass cover to rainfall variability: implications for the function and persistence of savanna
ecosystems. Adv Water Res. 28:291–302. Available from: <Go to ISI>://000227506500006
Scholes RJ, Dowty PR, Caylor K, Parsons DAB, Frost PGH, Shugart HH. 2002. Trends in
savanna structure and composition along an aridity gradient in the Kalahari. J Veg Sci.
13:419–428. Available from: http://www.jstor.org/stable/3236539
Scholes RJ, Walker BH. 1993. Cambridge studies in applied ecology and resource management.
Cambridge: Cambridge University Press; Xii + 306 p.
Swain H, Davis SM. 1978. Remote sensing: the quantitative approach. New York (NY):
McGraw- Hill.
Tews J, Blaum N, Jeltsch F. 2004. Structural and animal species diversity in arid and semi-arid
savannas of the southern Kalahari. Ann Arid Zone. 42:1–13.
Trimble. 2011a. Ecognition developer 8.0 reference book. Munchen, Germany.
Trimble. 2011b. Ecognition® developer 8.64. 1 user guide. Trimble Germany GmbH Munich,
Germany.
Turner MG. 2005. Landscape ecology in north america: past, present, and future. Ecology.
86:1967–1974.
 Geocarto International 877

Wiegand K, Saltz D, Ward D. 2006. A patch-dynamics approach to savanna dynamics and

woody plant encroachment – insights from an arid savanna. Perspect Plant Ecol Evol Syst.
7:229–242. Available from: http://www.sciencedirect.com/science/article/B7GVV-4HRDYDY-
1/2/81b524b9a1ffe4ce35ddaa60e344baa3
Wulder MA, Hall RJ, Coops NC, Franklin SE. 2009. High spatial resolution remotely sensed data
for ecosystem characterization. BioScience. 54:511–521. Available from: http://dx.doi.org/10.
1641/0006-3568(2004)054[0511:HSRRSD]2.0.CO;2
Yu Q, Gong P, Clinton N, Biging G, Kelly M, Schirokauer D. 2006. Object-based detailed vege-
tation classification with airborne high spatial resolution remote sensing imagery. Photogram-
metric Eng Remote Sens. 72:799–811. Available from: <Go to ISI>://WOS:0002387044
00010.
Yu Q, Gong P, Tian YQ, Pu R, Yang J. 2008. Factors affecting spatial variation of classification
uncertainty in an image object-based vegetation mapping. Photogrammetric Eng Remote
Sens. 74:1007–1018. Available from: <Go to ISI>://WOS:000258213100010
Downloaded by [University of Boras] at 06:38 05 October 2014