Journal of Applied Remote Sensing, Vol.

2, 023511 (20 March 2008)

Estimates of bare ground and vegetation cover from

Advanced Spaceborne Thermal Emission and
Reflection Radiometer (ASTER) short-wave-infrared
reflectance imagery

Tony K. Gill and Stuart R. Phinn

Centre for Remote Sensing and Spatial Information Science, The School of Geography,
Planning and Architecture, The University of Queensland, Brisbane QLD 4072, Australia.
t.gill1@[Link]

Abstract. The high level of success of estimating photosynthetic vegetation from

multispectral satellite sensors at regional scales has not been repeated for non-photosynthetic
vegetation and bare ground. Therefore regional scale estimates of total vegetation from
multispectral sensors are largely underestimated with implications for a wide range of
agricultural and environmental applications. Recent research using simulated data showed
that the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) had
the potential to provide reliable estimates of bare ground and total vegetation. This study built
on that research and found that estimates of bare ground retrieved from ASTER short-wave
infrared imagery using linear spectral unmixing correlated well with field measurements
(RMSE < 0.1, r2 > 0.7). Image endmember libraries required for spectral unmixing were
extracted from the image data using a combination of field knowledge and the lignin and
cellulose absorption index. The most reliable results were found by applying a sum-constraint
to the unmixing models and tying the signatures at wavebands that corresponded to cellulose
or clay-hydroxyl absorption features. The results of this research show that ASTER can
improve the estimates of total vegetation extracted from satellite imagery for environmental
studies at regional scales.
Keywords: bare ground, Lignin and Cellulose Absorption (LCA) index, Monte-Carlo
Spectral Mixture Analysis (MCSMA), vegetation cover.

1 INTRODUCTION
Measurements of tree and photosynthetically-active grass cover are acquired routinely using
satellite remote sensing [1-5]. The accurate retrieval of photosynthetic vegetation cover is due
to satellite sensors, such as Landsat ETM+, sampling the electromagnetic spectrum mostly in
the visible near-infrared (VNIR) region. In the VNIR region there are large differences in the
magnitude and shape between the photosynthetic and background spectral reflectance
signatures. However, the similarity of the spectral reflectance signatures of soil and non-
photosynthetic vegetation in the broad spectral bands (>10 nm) in the VNIR region, makes
their abundances difficult to estimate using multispectral sensors [6, 7]. Therefore, regional-
scale environmental monitoring and modelling applications may be underestimating total
vegetation cover when the estimates are derived from multispectral satellite observations in
the VNIR.
There are a wide range of applications that require reliable estimates of total vegetation
cover or, its reciprocal, the fraction of bare ground. As highlighted in previous research [6],
non-photosynthetic vegetation and organic litter are important in cropping and natural
ecosystems. For example, non-photosynthetic vegetation reduces water erosion, minimises
nutrient runoff, improves soil structure and fertility, and absorbs photosynthetically-active

© 2008 Society of Photo-Optical Instrumentation Engineers [DOI: 10.1117/1.2907748]

Received 9 Jan 2008; accepted 17 Mar 2008; published 20 Mar 2008 [CCC: 19313195/2008/$25.00]
Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 1

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 radiation that can lead to overestimates of biomass production if not accounted for. Therefore,
non-photosynthetic vegetation significantly impacts on the energy, nutrient, and carbon fluxes
in these systems [8, 9]. Knowledge of total biomass is also important for managing stocking
rates in grazing systems, and the distribution of total vegetation is an important indicator in
ecology for assessing food and shelter availability for insectivores, reptiles, and ground-
dwelling mammals.
Early research into quantifying the abundance of bare ground, non-photosynthetic
vegetation (NPV) and photosynthetic vegetation (PV) from remotely sensed systems used
airborne-hyperspectral instruments [6-8, 10-13]. Reliable estimates were obtained from the
airborne images using linear spectral unmixing in arid and semi-arid environments [10-12]
and cropping systems [13]. Vegetation indices, such as the cellulose absorption index (CAI),
have also been developed for estimating crop residues from hyperspectral imagery [6-8]. The
success of these methods can be attributed to distinct absorption features of the spectral
reflectance signatures for bare ground, NPV and PV in the shortwave-infrared (SWIR) region
between 2000nm and 2500nm (referred to in this paper as SWIR2). The absorption features
are due to clay-hydroxyl in soils at 2200nm [10, 14] and cellulose and lignin in vegetation
and plant litter near 2100nm and 2300nm [6, 8, 15]. The early research provided reliable
estimates of total vegetation at local scales (<10 km2). However, for many grazing, ecology
and climate applications these estimates are required at regional scales (10 km2 - 104 km2).
Recognising the need for estimates of total vegetation over larger areas, researchers
showed that bare ground and vegetation could be either discriminated using thermal infrared
imagery that is available from ASTER [16], or estimate their sub-pixel fractions using
simulated Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER)
SWIR data [7, 10]. However, there are a lack of similar studies that have used actual ASTER
imagery for the same purpose. ASTER is the most suitable spaceborne multispectral sensor
for such studies as it has five bands in the SWIR2 region with a 30m spatial resolution and a
wide spatial coverage (60km x 60km). Of the five bands in the SWIR2, three are particularly
useful for vegetation studies: band six because it samples the soil absorption feature at
2200nm; and bands five and eight because they sample the cellulose and lignin absorption
features near 2100nm and 2300nm [7]. If reliable estimates can be obtained using the limited
spectral sampling of ASTER, then the techniques may be useful with the next generation of
spaceborne hyperspectral instruments such as EnMap that sample more densely in the SWIR2
region [17].
A method that shows some promise in estimating the bare fractions from ASTER imagery
is linear spectral mixture analysis (SMA) [10]. Linear SMA models the pixel reflectance as a
linear combination of pure scene element (or endmember) reflectances. Linear SMA
calculates the contribution from each endmember signature to pixel reflectance and thus the
fractional abundance of each endmember. Optional constraints can be imposed on the model
so that the fractions sum to one and each fraction ranges from zero to one [18]. Tied
signatures, where the magnitude of the response at one band is subtracted from the others, can
be used instead of the original signatures [7, 10, 19]. With the correct tie point the
separability, and therefore variation, of the spectral reflectance signatures between the
endmembers can be improved. Monte-Carlo spectral mixture analysis (MCSMA) [10], which
is a variation of SMA, was used in this study since it accounts for the variation within an
endmember class. MCSMA also outputs a distribution (mean and standard deviation) of the
fractional estimates for each pixel that can be used to determine the quality of the unmixing
model and potentially the reliability of the extracted estimates.
A major challenge of SMA is identifying appropriate endmember signatures. When the
location of large pure targets is known, the endmember signatures can be extracted from the
image by matching it to the field sites. Alternatively, when field knowledge is unavailable,
automated or semi-automated endmember-extraction methods such as the pixel purity index
[20], N-FINDR [21], and others as summarised by Rogge et al [22] can be used. However, in

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 2

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 the absence of field knowledge, relating the automatically-extracted signatures to ground
features can be difficult. For vegetation studies, vegetation indices provide a method to
extract endmember signatures that can be related to ground features. For ASTER SWIR
imagery the lignin and cellulose absorption index (LCA) may be appropriate as it has been
found to be linearly related to non-photosynthetic residue cover [7].
In summary, there are many applications that would benefit from reliable satellite
estimates of bare ground and total vegetation cover at regional scales. The ASTER sensor, in
combination with the LCA index and linear spectral unmixing, has the potential to offer a
solution and needs to be tested. Therefore the overall objective of this study was to assess the
MCSMA method applied to ASTER SWIR2 imagery for estimating bare ground and
vegetation fractions. Firstly, the ability of the LCA index to identify bare and vegetation
endmembers from the image was investigated. Next, ASTER images were unmixed using
four combinations of mixture model constraints and three variations of the SWIR2 signatures
(SWIR2, SWIR2 tied at aster band five, SWIR2(5), and SWIR2 tied at band six, SWIR2(6)).
The image estimates were compared with field observations of fractional cover. Finally, an
analysis of the MCSMA statistics was undertaken to determine if they could provide an
alternative solution to field measurements for determining the accuracy of the image
estimates.

2 METHODOLOGY

2.1 Image and field data

A range of vegetation cover conditions was sampled on three field campaigns in Queensland,
Australia. Figure 1 shows the extent of three image acquisitions near the townships of
Charters Towers, Quilpie and Longreach. The Quilpie and Longreach sites were located
within semi-arid environments. Quilpie was dominated by sparse Acacia woodlands (Acacia
Cambagei, [Link]) interspersed with sparse eucalypt woodlands with short grasses and
open herblands. The Longreach site was dominated by tussock grasslands (Astrebla lappacea)
and Acacia open woodlands (Acacia Cambagei). The Charters Towers site was located on
savanna grasslands in a moderate rainfall climate (600-800mm annually) and was dominated
by eucalypt woodlands (Corymbia spp. and Eucalyptus spp) with an understorey of mixed
pastures [23, 24]. Photographs in Fig. 1 show the vegetation conditions typical of each study
region.
Sampling at Quilpie and Longreach took place at the end of the 2006 dry season (31
August to 3 September and 7-10 September respectively), and at Charters Towers at the
conclusion of the wet season in 2007 (29 April to 2 May). The time and location of the
sampling permitted a range of grass-cover conditions to be sampled, from low to medium
cover with little PV (Longreach and Quilpie), to medium to high cover with moderate
quantities of PV (Charters Towers). The Charters Towers site was the only study site with
cryptobiotic soil crusts that can cause confusion with PV due to similar spectral reflectance
properties when wet [25]. However, these crusts were not expected to impact significantly on
the results as they were dry at the time of sampling, only present at seven of the twelve plots,
and the field sampling showed that the exposed fraction was typically less than 10%.

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 3

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 Fig. 1. Location and extent of image acquisitions for the three field sites (boxes)
near the townships of Charters Towers, Longreach and Quilpie (triangles) with
annual average rainfall (left) and Landsat 7 ETM+ mosaic (right; False colour
composite with Red, NIR and Blue bands). Photographs that show the vegetation
typical of each study site are shown at the bottom.

At the Quilpie, Longreach and Charters Towers sites, twelve, fourteen and twelve plots
were sampled respectively. Each plot was 1 hectare (100 m x 100 m) in size and situated in
regions where the vegetation-cover pattern was regular. Each plot was sampled using three
intersecting 100 m transects. The first transect ran north-south, the other two transects were
aligned at 60 degrees and 120 degrees to the first transect, intersecting it at the 50 m point.
Plot centres were measured using a handheld GPS, with an estimated accuracy of less than
5m. The presence or absence of vegetation for three vertical layers (ground cover, midstorey
and overstorey) were noted at each metre along the transects. The results from each layer
were aggregated to derive the vertically-projected fractional-estimates of exposed bare ground
and vegetation. The color (value, hue and chroma) for the soil was also noted using Munsell
soil color charts. Spectral reflectance signatures of the vegetation at each plot were acquired
using a full range (350 nm - 2500 nm) Analytical Spectral Devices Field Spec Pro
spectrometer and laboratory-calibrated Spectralon® reflectance-panel. Spectral radiance
signatures of targets were obtained by directing the spectrometer fore-optic at a zenith angle
of 45 degrees in the plane perpendicular to the solar-plane at heights that minimised the
influence of background objects. Irradiance was measured using the Spectralon® panel and
used to convert the radiance signatures to reflectance.
ASTER images were acquired coincident with the field campaigns (Quilpie: 2 September
2006; Longreach: 9 September 2006; Charters Towers: 30 April 2007). ASTER is on board
NASA’s Terra satellite and contains three instruments that sample three bands in the VNIR, 6

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 4

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 bands in the SWIR, and 5 bands in the thermal-infrared. The SWIR imagery contains artifacts
due to light deflecting from the band 4 detector to other detectors in the SWIR instrument.
This phenomenon is known as cross-talk and affects bands 5 and 9 most severely; a
correction procedure is available [26, 27]. For this study, geometrically corrected, VNIR
surface reflectance and SWIR cross-talk-corrected surface reflectance images were acquired
from the Land Processes Distributed Active Archive Centre [28]. Comparison with field-
collected ground-control-points found that in the worst case the misalignment of the image
data was one pixel. Field sites were located in regions where the surrounding vegetation
pattern was consistent; therefore no further geometric corrections to the imagery were
required.

2.2 Estimating image fractions

2.2.1 Linear spectral unmixing

Linear SMA was used to derive sub-pixel fractions of bare ground and vegetation
endmembers from the image data. Linear SMA assumes that each photon reaching the sensor
has been reflected from only one endmember, of which shade might be one [12, 29]. Non-
linear mixing occurs when there is transmission of photons through the scene objects
resulting in multiple scattering with scene components [18]. In cases where transmittance and
scattering is high owing to complex vegetation canopies, the linear mixture model
incorporating shade may be a poor fit [12]. However, given that the endmembers in this study
were derived from the image, thereby implicitly capturing the shade component, and there is
limited transmittance of light through plant materials in the SWIR region, with a 30% worst
case for senescent materials [18], linear SMA was deemed to be a suitable method.
If the image contains m spectral bands and there are n endmembers, then the mixed
reflectance signature of an image pixel can represented by the matrix R~ = [~r1 ~r2 rm ] (T
~ T

denotes the transpose of the matrix), the endmembers can be stored as columns in the (m x n)
endmember matrix R , and the proportions of each endmember that contribute to the mixed
pixel can be represented the fractions matrix F = [ f1 f 2 f n ] . The linear mixture model
T

can be represented by the system of equations

~ (1)
R = RF + E

where E is the residual matrix [e1 e2 em ]T that represents the difference between the
~
observed and modelled pixel reflectance. Assuming that R and R are known, fractional
estimates for each endmember can be found by minimising E in Equation 1. In this study
Equation 1 was a well-defined system ( m ≥ n ) and solved using the QR-decomposition
routines available in the Lapack library [30]. The residual sum of squares (RSS) was
computed from E as a measure of the match between the modelled signature and the observed
pixel reflectance. The RSS was computed as

m m
RSS = ∑ (eλ ) = ∑ (~
rλ − Rλ F )
2 2
(2)
λ =1 λ =1

where Rλ is a row of R and represents the response of each endmember at wavelength λ.

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 5

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 2.2.2 Identifying image endmembers
Reference endmember signatures were extracted from the image data by matching them to
field sites with high fractions. For the Quilpie and Longreach sites those sample plots that had
high bare fractions (three plots at Quilpie: 99.33%, 91.67%, 89.67%; two plots at Longreach:
91.33%, 82.67%) were used to extract bare training-signatures from the image data. At
Charters Towers three sample plots with vegetation cover of 98.33%, 91.67% and 89.67%
were used to extract vegetation training-signatures. The spectral angle mapper [31] was used
to identify image endmembers where the spectral angle was less than 0.02 radians compared
to the training signatures.
Identifying sufficiently large sites dominated by vegetation at Quilpie and Longreach, and
bare ground at Charters Towers proved difficult. For these cases the LCA index was used to
identify image endmembers. The LCA index is specific to ASTER and provides a measure of
the relative depths of the lignin and cellulose absorption features and has been shown to be
linearly related to non-photosynthetic crop residue cover [7]. The LCA was computed as

LCA = 100[( ASTER 6 − ASTER 5) + ( ASTER 6 − ASTER 8)] . (3)

Comparison of the LCA index values with field measurements (Fig. 2) showed that a low
LCA index corresponded to low vegetation fractions (Quilpie and Longreach; LCA range
from -1 to 2) and higher LCA values to high vegetation fractions (Charters Towers; LCA
range from 1.5 to 3.5). High PV fractions decreased the correlations between the LCA index
and vegetation cover. At the Quilpie and Longreach sites, there was a reasonable linear
relationship of the LCA index with NPV cover (r2 of 0.395 and 0.443) because the PV
fractions were typically low (10 of 12 plots with less than 16.6% PV). However, the Charters
Towers relationship was poor due to high PV cover. The impact of high PV fractions on the
LCA index was consistent with previous findings [7]. Because the PV cover at Quilpie and
Longreach was low, pixels with the highest LCA values in the image (greater than 8 for
Quilpie and greater than 7 for Longreach) were used as vegetation endmember signatures. For
Charters Towers, pixels with the lowest LCA values in the image (less than -5) and where the
shape of the VNIR reflectance spectra did not resemble that of PV were used as bare ground
endmember-signatures.

Quilpie Longreach Charters Towers

0.5
2.5
1.5
3.0
0.4
2.0
1.0
PV fraction
LCA index

1.5 0.3
2.5
0.5
1.0 0.2
0.0 2.0
2
0.5 2 2 0.1
r =0.395 r =0.443 r =0.066
-0.5
0.0
1.5 0
0.00 0.10 0.20 0.1 0.3 0.5 0.4 0.5 0.6 0.7
Field Measured NPV

Fig. 2. Scatter plots of field-measured non-photosynthetic vegetation (NPV) fraction

versus the lignin and cellulose absorption (LCA) index from the ASTER images.
The color of the point indicates the photosynthetic (PV) fraction.

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 6

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 In many situations it is not possible to acquire field data at the same time the image is
acquired. Therefore the unmixing procedure was tested using endmember libraries that were
derived without field knowledge (referred to as image-image). To create the image-image
endmember libraries for Quilpie and Longreach, bare signatures were extracted where the
LCA index was less than -0.5 and -6 respectively and the shape of the VNIR reflectance
spectra did not match that of PV. For Charters Towers a second set of vegetation libraries
were extracted using pixels with a LCA index greater than 7.
Across all sites the total number of vegetation signatures ranged from 49 to 84. The
signatures were arranged into two distinct libraries (or bundles): bare ground and vegetation,
for use with MCSMA.

2.3 Monte Carlo spectral mixture analysis in the SWIR

MCSMA unmixes each pixel using one randomly selected endmember from each bundle. The
process is repeated multiple times for each pixel. The output is a distribution of fraction and
residual matrices (Equation 1) from which the mean and standard deviation of fractions and
RSS can be computed. The mean fraction is used as the final abundance estimate. MCSMA
provides the benefit of a statistical result without the need to unmix each pixel using every
combination of endmembers.
Equation 1 implies no constraints on the resulting fractions. However, constraints can be
applied by observing that fractions should range from 0 to 1 (referred to as endmember-
constrained; Equation 4) and in ideal cases the sum of the endmember contributions is 1
(referred to as sum-constrained; Equation 5). For this study all four possible combinations of
constraints were tested and are referred to by number (Table 1).

0 ≤ f i ≤ 1 , for i = 1,2, … , n (4)

∑f
i =1
i =1 (5)

The sum-constraint is straight-forward to implement by incorporating it into the linear

system of Equation 1. However solving endmember-constrained models is non-trivial and
alternative methods are required [18]. Using MCSMA, the endmember-constraint was
incorporated by accepting the results of unmixing runs where all output fractions were
between 0 and 1. Each pixel was unmixed until 50 valid results were found or a maximum
threshold of 500 unmixing runs was reached. An additional statistic referred to as the success
ratio, defined as the number of valid models to the total number of runs, was calculated for
each pixel.

Table 1. Spectral unmixing constraint-combinations tested in this study.

constraint description
1 Sum-constrained, endmember-constrained
2 Sum-constrained, endmember-unconstrained
3 Sum-unconstrained, endmember-constrained
4 Sum-unconstrained, endmember-unconstrained

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 7

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 Bands 5 and 6 were selected as the tie points due to vegetation and soil-related absorption
features that correspond with these bands. The resultant signatures are referred to as
SWIR2(5) and SWIR2(6) respectively. In total, 24 variants of the unmixing model were
applied to each study site derived from combining 3 tie points, 4 constraints, and 2 different
sets of endmember bundles (field-image and image-image).

2.4 Validation of the image fractions

The statistics ouput from spectral unmixing can be used to assess how well the model
reproduces the observed reflectance. However, this is different from determining how well
the derived fractions match those on the ground. In this study the image-derived fractions
were compared to the field measured fractions using regression analysis. In addition, the
MCSMA statistics were analysed to determine if they were useful indicators of the accuracy
of the derived fractional estimates.

3 RESULTS AND DISCUSSION

3.1 General trends

Table 3 shows that the sum-unconstrained unmixing models gave better estimates of bare
fractions than vegetation cover. Also, using field-image bundles consistently produced better
results than using image-image bundles. A possible reason is that the endmember-bundle
signatures derived using field knowledge were a better match for the scene vegetation than
those extracted using vegetation indices alone. Field campaigns, coincident with image
acquisitions are therefore required when using the method outlined in this paper. For the rest
of the discussion, only estimates of bare fractions derived using the field-image endmember-
bundle combinations are considered.

3.2 The impact of endmember selection on results

The fraction images in Fig. 3 show the overall trend of high bare fractions at Quilpie, medium
to high at Longreach and low at Charters Towers. Landscape features are also clearly visible.
Images 3-1a and 3-1b show clear separation between the dense vegetation and river beds
within the channel systems. An unsealed road in image 3-2a can be seen to dissect a large
patch of woodland that is bordered by a region of low vegetation cover to the north and west.
A grazing paddock with high vegetation cover relative to the neighbouring paddocks is
centred within image 3-2b. Similarly, differences in vegetation cover between grazing
paddocks are visible in image 2-3a, although the total cover is much higher compared with 3-
2b. Finally, the VNIR scene for image 3-3b shows a natural boundary due to a basalt flow.
However, the boundary is not as clear in the corresponding bare-fractions image. The actual
vegetation cover across the boundary is similar and suggests that the unmixing method used
in this study is invariant to soil color. While Fig. 3 provides a good visual validation of the
method, the discussion below provides a more detailed analysis of the results.
The mean field-collected spectral signatures for the three study sites are shown in Fig. 4
and their standard deviations in Table 2. The absorption features for soils at 2200 nm, and
broad absorption features for NPV near 2100 nm are consistent with results from the literature
[7, 10]. The signatures convolved to the ASTER bands (bottom of Fig. 4) also highlight the
absorption feature of soils at 2200 nm. The endmembers extracted from the image (Fig. 5)
contained the same general features as the field spectra convolved to the ASTER bands, with
the exception of band 9 (2395 nm). The features were the absorption at 2200 nm for soil
spectra and absorption at 2000 nm for cellulose and lignin. These features are consistent with
previous findings [7, 10, 14]. The increased reflectance at band 9 for the image signatures was
most likely due to incomplete correction for cross-talk at that band [27].

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 8

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 Fig. 3. Bare-fraction estimates calibrated to field observations for selected regions of
each study site (right) shown alongside the ASTER VNIR reflectance imagery (left).

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 9

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 Fig. 4. Top: Mean field-collected spectral reflectance signatures in the SWIR2
region with the ASTER spectral response functions for bands 5 to 9. Bottom: the
signatures convolved to the ASTER bands; the vertical lines show position of bands
5, 6 and 8 used to compute the lignin and cellulose absorption (LCA) index.

Table 2. The standard deviation of the ASTER-convolved field collected spectra for
Quilpie (Q), Longreach (L) and Charters Towers (C).
Band 5 6 7 8 9
Q 0.11 0.103 0.105 0.105 0.103
L 0.099 0.096 0.095 0.098 0.097
Bare

C 0.083 0.079 0.085 0.082 0.080

Q 0.185 0.188 0.178 0.171 0.170
L 0.076 0.076 0.072 0.069 0.068
NPV

C 0.101 0.102 0.097 0.096 0.097

Q 0.056 0.056 0.053 0.048 0.046
L 0.054 0.054 0.050 0.046 0.043
PV

C 0.036 0.037 0.033 0.029 0.027

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 10

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 0.5 Quilpie Longreach Charters Tow ers

0.4

SWIR2
0.3

Reflectance/Tied Reflectance
0.2

0.1

0.075

SWIR2(5) 0.025

-0.025

-0.075

0.075 bare
SWIR2(6)

veg
0.025

-0.025

-0.075
2.2 2.3 2.4 2.2 2.3 2.4 2.2 2.3 2.4
Wavelength (µm)
Fig. 5. The mean and standard deviation (error bars) for SWIR2, SWIR2 tied at
band 5 (SWIR2(5)), and SWIR2 tied at band 6 (SWIR2(6)) spectral signatures of
the image endmember-bundles for the three study sites.

Table 3 shows the unmixing models that had reasonable agreement with the field
measured bare fractions (r2 > 0.5 and RMSE of near 0.2 or less). The RMSE and r2 statistics
were computed using 5-fold cross validation [32] with outliers removed. An outlier was
defined as an image fraction that was more than 0.2 from the regression line. Figure 6 shows
the statistics and regression plots for the best unmixing models for each site. Some of the
models for Longreach and Charters Towers sites generated outlying data points, and Table 3
also contains the results of the models once outliers in the bare estimates were removed (plot
3 for Quilpie SWIR2(6), plot 8 for Charter Towers SWIR2(5) and plot 3 for Charters Towers
SWIR2). With outliers removed, the regression relationships in Table 3 showed that the
Charters Towers bare fractions were slightly underestimated, the Quilpie estimates varied
about the 1:1 line, and the Longreach estimates were overestimated. These results reflect the
mixed nature of the field plots used to determine the original reflectance spectra. For
Longreach the training plots for bare ground were likely to contain some contribution from
vegetation (one plot having 18.33% cover). Therefore more contribution from soil spectra in
the unmixing models was required to account for the extra contribution from vegetation in the
bare signatures. For Charters Towers the training plots for vegetation contained bare fractions
(one plot being 12% bare), therefore more contribution from vegetation spectra in the
unmixing models was required to offset the bare signal. Finally, for Quilpie, the bare plots
had very high fractions (one plot estimated at 99.33%), therefore signatures were more likely
to be pure and hence the models showed no clear trend to under or overestimate the bare
fractions.

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 11

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 Table 3. Results from unmixing models that gave good results for bare fractions using the SWIR2,
SWIR2 tied at band 5 (SWIR2(5)), and SWIR2 tied at band 6 (SWIR2(6)) spectral signatures. A *
indicates the well-performed models for each site, a + indicates model performance after outlying plots
were removed from the regression. The left hand side of the regression equation (y) refers to the image
estimates.
Bare Bare Bare Veg Veg
plot constraint bands (tie) RMSE r2 Regression RMSE r2
Quil* 2 SWIR2(5) 0.101 0.530 y=0.915x+0.113 0.101 0.577
Quil 3 SWIR2(5) 0.111 0.595 y=1.044x-0.086 0.135 0.033
Quil 4 SWIR2(5) 0.130 0.716 y=1.552x-0.416 0.131 0.331
Quil 3 SWIR2 0.093 0.514 y=0.812x+0.125 0.122 0.191
Quil 4 SWIR2 0.104 0.573 y=1.101x-0.069 0.121 0.370
Quil 2 SWIR2(6) 0.131 0.522 y=1.039x+0.017 0.129 0.500
Quil 4 SWIR2(6) 0.123 0.562 y=1.362x-0.249 0.135 0.315
Long 1 SWIR2(6) 0.196 0.510 y=0.770x+0.329 0.194 0.464
Long+ 1 SWIR2(6) 0.169 0.795 y=1.111x+0.089 NA NA
Long 2 SWIR2(6) 0.203 0.665 y=1.034x+0.167 0.205 0.662
Long* 2 SWIR2(6) 0.185 0.852 y=1.297x-0.027 NA NA
Chat 2 SWIR2(5) 0.139 0.550 y=1.007x-0.097 0.148 0.563
Chat* 2 SWIR2(5) 0.140 0.721 y=1.123x-0.140 NA NA
Chat 1 SWIR2 0.114 0.553 y=1.135x-0.019 0.110 0.551
Chat+ 1 SWIR2 0.062 0.770 y=0.928x-0.011 NA NA
Chat 2 SWIR2 0.144 0.557 y=1.372x-0.095 0.141 0.541
Chat+ 2 SWIR2 0.108 0.678 y=1.152x-0.086 NA NA
Chat 1 SWIR2(6) 0.080 0.790 y=0.532x+0.040 0.079 0.790
Chat 3 SWIR2(6) 0.101 0.640 y=0.363x+0.060 0.180 0.085

Soil color is another factor that may cause bare fractions to be poorly estimated.
Longreach and Charters Towers had the greatest variation in soil color across the sample
plots. Figure 6 shows the outlying plot for Charters Towers had a brighter and redder soil than
the other plots. However the soil color for the outlier at Longreach was similar to the other
plots so there was no clear evidence that outlying estimates are affected by soil color. The
Quilpie study site had the least variation in soil colors, yet the lowest r2 with field estimates.
In summary, this study was unable to establish a clear trend between soil color and the
resulting fractions. If there was some variation in the results caused by soil color it was small
compared to the variation caused by endmember selection.
The image-derived estimates of vegetation using sum-unconstrained unmixing did not
match the field observations as well as those for bare ground as shown by the poor r2 values
in Table 3. The poor vegetation estimates at the Quilpie and Longreach sites were likely to be
due to the high variability of the spectral signatures within the vegetation bundles (Fig. 5).
For Charters Towers the spectral variability within the vegetation bundle was low, however
the spectral variation of vegetation in the scene was high due to a large number of vegetation
species and mixture of photosynthetic and non-photosynthetic grasses. Future studies may
include two or more vegetation endmembers to capture the variation across the scene and
reduce the variation within each bundle. Minimising the spectral variability within bundles
should improve the performance of the MCSMA algorithm as shown in previous research [7,
10, 19].

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 12

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 1.25 Quilpie
SWIR2(5)
1.1

0.95

0.8
RMSE:=0.100
0.65 2
r =0.547
y=0.915x + 0.113
0.5
0.5 0.6 0.7 0.8 0.9 1

1.25 Longreach
SWIR2(6)
1.1
Image Fraction

0.95 regression
1:1 line
0.8 outlier

0.65
RMSE:=0.184
2
r =0.851
0.5
y=1.295x + -0.027
0.5 0.6 0.7 0.8 0.9 1

0.5 Charters Towers

SWIR2(5)
0.35

0.2

0.05
RMSE:=0.139
-0.1 2
r =0.690
-0.25
y=1.123x + -0.140
0 0.1 0.2 0.3 0.4 0.5

Field Fraction
Fig. 6. The mean (points) and standard deviation (error bars) of the bare fractions
derived from image data using Monte-Carlo spectral mixture analysis with sum-
constrained unmixing and spectral signatures tied at ASTER band 5 (SWIR2(5)) and
ASTER band 6 (SWIR2(6)). The color of the points corresponds to the soil color
observed in the field.

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 13

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 The final permutation on endmember signature selection was the choice of tie point.
Across all three study sites the SWIR2(6) signatures were found to produce estimates of bare
fractions that had reasonable agreement with field observations, and the only signatures that
produced reasonable results for the Longreach site. Figure 5 shows that for Longreach the
endmember classes were more spectrally-distinct when tied at band 6, compared to band 5 or
untied signatures, and may explain the better results obtained by tying at band 6 at this site.
However, SWIR2(5) and SWIR2(6) signatures produced reasonable results for Quilpie
despite the spectral separability of the SWIR2(5) signatures being similar to that of the
SWIR2(5) signatures at Longreach. Therefore, the impact of the tie point on the results in
these cases was not clear. SWIR2(5) signatures were most appropriate for unmixing at the
Quilpie and Charters Towers study sites with SWIR2 signatures also producing good results.
In general, tying the ASTER signatures at bands 5 or 6 produced slightly better estimates of
bare fractions than SWIR2 alone and is in agreement with previous findings [10]. The final
choice of tie point appears to be specific to the study area and requires comparison with field
measured estimates.

3.3 Analysis of the unmixing constraints

The constraint combinations that were deemed best for estimating the bare fractions at each
site are indicated in Table 3. The sum-unconstrained models were more successful at the
Quilpie site while the sum-constrained models were found to produce reliable bare fractions
at all sites. Similarly the endmember-unconstrained models were the most reliable at the
Quilpie site, and just as reliable as the endmember-constrained models at the other sites. As
discussed in the section 3.2, the bare endmembers extracted from the Quilpie image were
purer than the bare endmembers for Longreach and vegetation endmembers for Charters
Towers with the consequence that constraints were not required to produce reliable estimates.
When the RMSE, r2, and regression slope were considered, constraint 2 was deemed to be the
most consistent across all sites.
Careful choice of constraints must be made if the purpose of the study is to calibrate the
unmixed image with field estimates. Endmember-constrained models force extreme estimates
to be either zero or one and are therefore biased. Endmember-unconstrained models were
considered to be more appropriate in this study because they are unbiased at the extremes.
Therefore the regression relationships of the constraint 2 models were used to calibrate the
unmixed images to field-measured fractions to produce the fraction images shown in Fig. 3.
The small number of sample plots at each site was insufficient to provide a robust analysis of
the accuracy of the calibrated images.

3.4 Analysis of the MCSMA statistics

One standard deviation of the bare fractions (error bars in Fig. 6), were derived from the
MCSMA algorithm, and were found to be lower at the Longreach site compared to the other
two sites. Fig. 5 shows that the variation in the bare signatures used in the unmixing models
for Longreach was also much lower than the other sites suggesting a relationship between the
variation of the endmember bundles and the variation of the corresponding fractional
estimates. However, the variation in the vegetation signatures used to unmix the Charters
Towers image was also low, but the resultant variation in bare fractions higher than for
Longreach. Analysis of the variation of the field-collected spectra (Table 2) shows that the
NPV signatures at Quilpie and Charters Towers had greater spectral variability than at
Longreach. Hence the variation in the image fractions was related to the variation in the
endmember bundle libraries, the spectral variability of the targets in the scene, and how well
the libraries matched the signatures of the scene targets. Improving the derived fractions
requires changes to the MCSMA algorithm to constrain endmember selection on a per-pixel

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 14

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 basis. Such constraints include building the endmember bundles based on spatial proximity to
the pixel [22] or stratifying the image and spectral libraries by land type.
The standard deviation of the estimates appear to be unrelated to the closeness of the
match between the image and field-measured fractions. The estimates at Longreach, while
having low variability, also exhibited a high RMSE (Fig. 6). The opposite was true for the
Quilpie site. Also, Fig. 6 shows that the standard deviations of the fractional estimates for the
outlying plots were similar to the other plots. Therefore the output standard deviations only
indicated how well the modelled reflectance matched the pixel signature.
Examination of the mean and standard deviation of the residual sum of squares (MRSS
and SDRSS; Table 4) showed that the statistics for the outlying plots were similar to the other
plots at the same site. Also, plot 10 at Charters Towers and plot 5 at Longreach had high
MRSS statistics, but the final fractional estimates were not considered outliers. Therefore, the
MCSMA statistics provided information about the match between the modelled and pixel
signatures, but could not be used to indicate how well the image estimates matched the field
measurements.
Finally, the success ratio can be considered an indicator of model fit, but needs to be
interpreted in context with the unmixing constraints and fraction of ground cover. If the
unmixing model is endmember-unconstrained a valid model is found for every run and the
success ratio will always be 1. In such cases the MRSS, SDRSS and fraction standard-
deviations are the indicators of model fit. For endmember-constrained unmixing, low success
ratios are often associated with extreme fractions. Table 5 shows low success ratios for high
bare fractions at Quilpie (plot 5) and Longreach (plots 5 and 14), and high vegetation
fractions at Charters Towers (plots 7 and 10). In these cases estimates outside the range 0 to 1
were rejected regardless of whether they were close to the actual fractions; therefore the
success ratio could not be used as an indicator of model fit. In non-extreme cases the success-
ratio could be used as an indicator for how well the unmixing model matches the pixel
signature. For example, Table 5 shows that plot 10 for Charters Towers had a low success
ratio compared to other plots at the same site with similar bare fractions, and the result must
be considered highly uncertain.

Table 4. Mean residual sum of squares and standard deviation of the residual sum of squares (MRSS ±
SDRSS) for spectral unmixing models shown in Fig. 6. * indicates outlying plots.

Plot # Quil Long Chat

1 0.017±0.007 0.005±0.001* 0.013±0.005
2 0.008±0.005 0.002±0.001 0.008±0.005
3 0.01±0.005 0.012±0.003 0.016±0.007
4 0.005±0.006 0.007±0.001 0.012±0.004
5 0.017±0.01 0.018±0.001 0.009±0.006
6 0.006±0.004 0.004±0.001 0.013±0.007
7 0.011±0.005 0.005±0.001 0.012±0.007
8 0.019±0.007 0.009±0.001 0.013±0.009*
9 0.008±0.005 0.006±0.001 0.009±0.006
10 0.012±0.005 0.007±0.002 0.024±0.012
11 0.018±0.006 0.006±0.001 0.012±0.006
12 0.021±0.008 0.008±0.001 0.018±0.007
13 0.012±0.001
14 0.01±0.001

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 15

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 Table 5. The field measured bare fractions and unmixing model success ratios for selected endmember-
constrained models: Quilpie, constraint 3, SWIR2(5); Longreach, constraint 1 SWIR2(6); Charters
Towers, constraint 1 SWIR2.

Quil Long Chat

Plot # frac [Link] frac [Link] frac [Link]
1 0.786 0.980 0.533 1 0.12 0.943
2 0.777 0.877 0.667 1 0.17 0.980
3 0.803 0.526 0.57 1 0.273 0.943
4 0.897 0.746 0.637 1 0.31 0.980
5 0.993 0.199 0.75 0.146 0.197 0.980
6 0.817 0.926 0.657 1 0.46 1
7 0.917 0.214 0.7 1 0.083 0.370
8 0.743 0.667 0.657 1 0.12 0.909
9 0.697 0.943 0.63 1 0.183 0.515
10 0.653 0.735 0.49 1 0.013 0.048
11 0.863 0.820 0.914 0 0.173 0.980
12 0.553 0.926 0.73 1 0.18 0.962
13 0.617 1
14 0.827 0.4

4 CONCLUSIONS
This research produced reliable estimates of bare fractions by linear spectral unmixing of
ASTER SWIR imagery in two semi-arid and one savanna environment in Australia. The
method employed the MCSMA algorithm and found that obtaining consistent results across
all sites was possible by applying a sum-constraint and tying the spectral signatures at
ASTER bands 5 and 6. Further improvement of the estimates requires calibration of the
image estimates to ground observed fractions. Therefore, field work, coincident with image
acquisition, is an essential part of the method.
A combination of field knowledge and the LCA index was used to identify image
endmembers. The ability of the LCA index to identify pure pixels was dependent on whether
there were indeed pure pixels in the scene. The lack of pixels with high vegetation in the
Quilpie and Longreach scenes, and lack of pixels with high bare fractions in the Charters
Towers scene resulted in poor fractional estimates when vegetation indices were used to
identify both sets of endmember bundles. Significant improvements in the bare estimates
were made when one of the endmember bundles was derived using field knowledge.
The variation of vegetation within the scene appears to have impacted on the ability of the
method to directly estimate the vegetation fraction. Instead, vegetation can be estimated as the
reciprocal of bare ground. Large variation in spectral-reflectance signatures within each
endmember-bundle resulted in bare-ground fractions that had high standard deviations as
computed from MCSMA. Of the MCSMA statistics investigated, all were found to be good
indicators of how well a modelled signature matched the image pixel. However, none of the
MCSMA statistics were found to be indicators of how well the estimated fractions matched
the field measurements. The only method that was reliable was direct comparison with
ground measured fractions.
In summary, reliable estimates of bare fractions and therefore total vegetation can be
obtained using ASTER SWIR imagery, the MCSMA algorithm and field data. Therefore,
regional-scale estimates of total vegetation and bare ground can be made available for a wide

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 16

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 range of grazing, ecology and environmental applications. However, improvements can still
be made by reducing field-work requirements and partitioning total vegetation into
contributions from the photosynthetic and non-photosynthetic components. The results also
indicate that future land-observing satellite sensors that include multiple bands in the 2000-
2500 nm range will improve vegetation cover estimates. For example, the upcoming
hyperspectral satellite sensor Enmap, will sample in the SWIR2 region [17] and has the
potential to greatly improve estimates of the major land surface components.

Acknowledgments
This research was funded through an Australian Postgraduate Award and the CSIRO Division
of Exploration and Mining. The authors wish to thank Dr. Terry Beutel of the Department of
Primary Industries for assisting with the cost of image purchase and linking us with graziers
in Western Queensland. We are grateful to the graziers for permitting access to their
properties and providing us with accommodation. And finally to Nicholas Webb for field
assistance, the ASTER science team for scheduling image acquisitions, the Australian Bureau
of Meteorology for providing the rainfall map in Fig. 1, and Dr. Peter Scarth who commented
on an early draft of the manuscript.

References

[1] G. Pickup, V. H. Chewings, and D. J. Nelson, "Estimating changes in vegetation

cover over time in arid rangelands using landsat MSS data," Rem. Sens. Environ. 43,
243-263 (1993) [doi:10.1016/0034-4257(93)90069-A].
[2] R. P. Pech, R. D. Graetz, and A. W. Davis, "Reflectance Modeling and the
Derivation of Vegetation Indexes for an Australian Semiarid Shrubland," Int. J. Rem.
Sens. 7, 389-403 (1986) [doi:10.1080/01431168608954694].
[3] T. J. Danaher, J. D. Armston, and L. J. Collett, "A multiple regression model for the
estimation of woody foliage cover using Landsat in Queensland, Australia,"
presented at Int. Geosci. Rem. Sens. Symp., Anchorage, (2004).
[4] R. D. Graetz, R. P. Pech, and A. W. Davis, "The Assessment and Monitoring of
Sparsely Vegetated Rangelands Using Calibrated Landsat Data," Int. J. Rem. Sens.
9, 1201-1222 (1988) [doi:10.1080/01431168808954929].
[5] H. Lu, M. R. Raupach, T. R. McVicar, and D. J. Barrett, "Decomposition of
vegetation cover into woody and herbaceous components using AVHRR NDVI time
series," Rem. Sens. Environ. 86, 1-18 (2003) [doi:10.1016/S0034-4257(03)00054-3].
[6] P. L. Nagler, Y. Inoue, E. P. Glenn, A. L. Russ, and C. S. T. Daughtry, "Cellulose
absorption index (CAI) to quantify mixed soil-plant litter scenes," Rem. Sens.
Environ. 87, 310-325 (2003) [doi:10.1016/[Link].2003.06.001].
[7] C. S. T. Daughtry, E. R. Hunt, P. C. Doraiswamy, and J. E. McMurtrey, "Remote
sensing the spatial distribution of crop residues," Agronomy J. 97, 864-871 (2005)
[doi:10.2134/agronj2003.0291].
[8] C. S. T. Daughtry, Hunt E.R., McMurtrey III J.E., "Assessing crop residue cover
using shortwave infrared reflectance," Rem. Sens. Environ. 90, 126-134 (2004)
[doi:10.1016/[Link].2003.10.023].
[9] P. L. Nagler, C. S. T. Daughtry, and S. N. Goward, "Plant litter and soil reflectance,"
Rem. Sens. Environ. 71, 207-215 (2000) [doi:10.1016/S0034-4257(99)00082-6].
[10] G. P. Asner and K. B. Heidebrecht, "Spectral unmixing of vegetation, soil and dry
carbon cover in arid regions: comparing multispectral and hyperspectral
observations," Int. J. Rem. Sens. 23, 3939-3958 (2002)
[doi:10.1080/01431160110115960].

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 17

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 [11] G. S. Okin, D. A. Roberts, B. Murray, and W. J. Okin, "Practical limits on
hyperspectral vegetation discrimination in arid and semiarid environments," Rem.
Sens. Environ. 77, 212-225 (2001) [doi:10.1016/S0034-4257(01)00207-3].
[12] D. A. Roberts, M. O. Smith, and J. B. Adams, "Green vegetation, nonphotosynthetic
vegetation, and soils in AVIRIS data," Rem. Sens. Environ. 44, 255-269 (1993)
[doi:10.1016/0034-4257(93)90020-X].
[13] A. Bannari, A. PacheCo, K. Staenz, H. McNairn, and K. Omari, "Estimating and
mapping crop residues cover on agricultural lands using hyperspectral and IKONOS
data," Rem. Sens. Environ. 104, 447-459 (2006) [doi:10.1016/[Link].2006.05.018].
[14] J. R. Irons, R. A. Weismiller, and G. W. Peterson, "Soil Reflectance," in Theory and
Applications of Optical Remote Sensing, G. Asrar, Ed., John Wiley and Sons, New
York (1989).
[15] C. D. Elvidge, "Visible and near infrared reflectance characteristics of dry plant
materials," Int. J. Rem. Sens. 11, 1775-1795 (1990)
[doi:10.1080/01431169008955129].
[16] A. N. French, T. J. Schmugge, and W. P. Kustas, "Discrimination of senescent
vegetation using thermal emissivity contrast," Rem. Sens. Environ. 74, 249-254
(2000) [doi:10.1016/S0034-4257(00)00115-2].
[17] T. Stuffler, C. Kaufmann, S. Hofer, K. P. Forster, G. Schreier, A. Mueller, A.
Eckardt, H. Bach, B. Penne, U. Benz, and R. Haydn, "The EnMAP hyperspectral
imager- An advanced optical payload for future applications in Earth observation
programmes," Acta Astronautica 61, 115-120 (2007)
[doi:10.1016/[Link].2007.01.033].
[18] N. A. Drake, S. Mackin, and J. J. Settle, "Mapping Vegetation, Soils, and Geology in
Semiarid Shrublands Using Spectral Matching and Mixture Modeling of SWIR
AVIRIS Imagery," Rem. Sens. Environ. 68, 12-25 (1999) [doi:10.1016/S0034-
4257(98)00097-2].
[19] G. P. Asner and D. B. Lobell, "A biogeophysical approach for automated SWIR
unmixing of soils and vegetation," Rem. Sens. Environ. 74, 99-112 (2000)
[doi:10.1016/S0034-4257(00)00126-7].
[20] J. W. Boardman, "Automating spectral unmixing of AVIRIS data using convex
geometry concepts," presented at Summaries 4th Annual JPL Airborne Geosci.
Worksh., JPL Publication 93-26, (1993).
[21] M. E. Winter, "Fast autonomous spectral endmember determination in hyperspectral
data.," presented at Proc. 13th Int. Conf. Applied Geologic Rem. Sens., Vancouver,
B.C. (1999).
[22] D. M. Rogge, B. Rivard, J. Zhang, A. Sanchez, J. Harris, and J. Feng, "Integration of
spatial-spectral information for the improved extraction of endmembers," Rem. Sens.
Environ. 110, 287-303 (2007) [doi:10.1016/[Link].2007.02.019].
[23] Department of Environment and Water Resources (2007). National Vegetation
Information System. Online. URL: [Link]
[accessed: 20/6/2006]
[24] Environmental Protection Agency (2007). Regional Ecosystem Description Database
(REDD). Online. URL:
[Link]
[accessed: 6/6/2006]
[25] A. Karnieli, G. J. Kidron, C. Glaesser, and E. Ben-Dor, "Spectral Characteristics of
Cyanobacteria Soil Crust in Semiarid Environments," Rem. Sens. Environ. 69, 67-75
(1999)[doi:10.1016/S0034-4257(98)00110-2].
[26] A. Iwasaki, H. Fujisada, H. Akao, O. Shindou, and S. Akagi, "Enhancement of
Spectral Separation Performance for ASTER/SWIR," Proc. SPIE. 4486, 42-50
(2002)[doi:10.1117/12.455140].

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 18

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]

 [27] A. Iwasaki and H. Tonooka, "Validation of a crosstalk correction algorithm for
ASTER/SWIR.," IEEE Trans. Geosci. Rem. Sens. 43, 2747-2751 (2005).
[28] NASA (2006). Land Processes Distributed Archive Center. Online. URL:
[Link] [accessed: 22/4/2006]
[29] M. O. Smith, S. L. Ustin, J. B. Adams, and A. R. Gillespie, "Vegetation in deserts: I.
A regional measure of abundance from multispectral images," Rem. Sens. Environ.
31, 1-26 (1990) [doi:10.1016/0034-4257(90)90074-V].
[30] E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. Dongarra, J. Du Croz,
A. Greenbaum, S. Hammarling, A. McKenney, and D. Sorensen, LAPACK Users'
Guide, 3rd ed., Society for Industrial and Applied Mathematics, Philidelphia (1999).
[31] F. A. Kruse, A. B. Lefkoff, J. W. Boardman, K. B. Heidebrecht, A. T. Shapiro, P. J.
Barloon, and A. F. H. Goetz, "The spectral image processing system (SIPS)--
interactive visualization and analysis of imaging spectrometer data," Rem. Sens.
Environ. 44, 145-163 (1993) [doi:10.1016/0034-4257(93)90013-N].
[32] B. Efron and R. Tibshirani, J., "Cross-validation and other estimates of prediction
error," in An introduction to the Bootstrap, vol. 57, Monographs on Statistics and
Applied Probability, pp. 239-241, Chapman Hall, New York (1993).

Journal of Applied Remote Sensing, Vol. 2, 023511 (2008) Page 19

Downloaded From: [Link] on 10/13/2015 Terms of Use: [Link]