Spectral variability within species and its effects on Savanna tree species discrimination

Abstract

Differences in within-species phenology and structure driven by factors including topography, edaphic properties, and climatic variables present important challenges for species differentiation with remote sensing in the Kruger National Park, South Africa. The objective of this study was to examine probable factors including intraspecies spectral variability and the spectral sample size that could affect remote sensing of Savanna tree species across a land-use gradient in the Kruger National park. Eighteen species were examined: Acacia gerradii, Acacia nigrescens, Combretum apiculatum, Combretum collinum, Combretum hereroense, Combretum imberbe, Combretum zeyheri, Dichrostachys cinerea, Euclea sp (E. divinurum and E. natalensis, Gymnosporia sp (G. buxifolia and G. senegalensis), Lonchocarpus capassa, Peltoforum africanum, Piliostigma thonningii, Pterocarpus rotundifolia, Sclerocarya birrea, Strychnos sp (S. madagascariensis, S. usambarensis), Terminalia sericea and Ziziphus mucronata. Discriminating species using the K-nearest neighbour (K = 1) classifier with spectral angle mapper (SAM) yielded a higher classification accuracy (48% overall accuracy) compared to 16% for the classification involving the mean spectra for each species as the training spectral set. Within-species spectral variability and the training sample size were identified as important factors affecting classification accuracy of the tree species. We recommend a non-parametric classifier such as K-nearest neighbour classifier for classifying and mapping tree species in a highly complex environment such as the savanna system of the Kruger National Park.

Full text

Page 1

SPECTRAL VARIABILITY WITHIN SPECIES AND ITS EFFECTS ON SAVANNA TREE
SPECIES DISCRIMINATION

Moses A. Cho1, Pravesh Debba2, Renaud Mathieu1, Jan van Aardt3, Greg Asner4, Laven Naidoo1,
Russell Main1, Abel Ramoelo1, Bongani Majeke1,.

1Council for Scientific and Industrial Research (CSIR), Natural Resources and the Environment, Ecosystems- Earth
Observation, P.O. Box 395, Pretoria, 0001, South Africa.
2CSIR, Built Environment, Box 395, Pretoria, South Africa
3 RIT: Center for Imaging Science, Laboratory for Imaging Algorithms and Systems
54 Lomb Memorial Drive, Building 17-3173, Rochester NY 14623, USA
4Department of Global Ecology, Stanford University, 260 Panama St., Stanford CA 94305, USA


ABSTRACT

Differences in within-species phenology and structure driven by
factors including topography, edaphic properties, and climatic
variables present important challenges for species differentiation
with remote sensing in the Kruger National Park, South Africa.
The objective of this study was to examine probable factors
including intraspecies spectral variability and the spectral sample
size that could affect remote sensing of Savanna tree species across
a land-use gradient in the Kruger National park. Eighteen species
were examined: Acacia gerradii, Acacia nigrescens, Combretum
apiculatum, Combretum collinum, Combretum hereroense,
Combretum imberbe, Combretum zeyheri, Dichrostachys cinerea,
Euclea sp (E. divinurum and E. natalensis, Gymnosporia sp (G.
buxifolia and G. senegalensis), Lonchocarpus capassa, Peltoforum
africanum, Piliostigma thonningii, Pterocarpus rotundifolia,
Sclerocarya birrea, Strychnos sp (S. madagascariensis, S.
usambarensis), Terminalia sericea and Ziziphus mucronata.
Discriminating species using the K-nearest neighbour (K = 1)
classifier with spectral angle mapper (SAM) yielded a higher
classification accuracy (48% overall accuracy) compared to 16%
for the classification involving the mean spectra for each species as
the training spectral set. Within-species spectral variability and the
training sample size were identified as important factors affecting
classification accuracy of the tree species. We recommend a non-
parametric classifier such as K-nearest neighbour classifier for
classifying and mapping tree species in a highly complex
environment such as the savanna system of the Kruger National
Park.

Index Terms— Savanna tree species; Spectral variability;
Multiple endmember approach; Spectral angle mapper

1. INTRODUCTION

The ability to map vegetation to the species level is of wide interest
in Ecology. Species-level maps of vegetation have important
applications in resource inventories, biodiversity assessment, and
fire hazard assessment. Species mapping with remote sensing is
based on the assumption that each species has a unique spectral
signature. Spectral signatures of vegetation vary according to
biochemical content, physical structure of plant tissues and canopy
architecture.

Several mapping methods are applied in remote sensing to quantify
species or vegetation community distribution at the local to
regional scale. The most commonly used methods include
maximum likelihood, spectral mixture analysis (SMA)[1] and
spectral angle mapper (SAM)[2]. The application of some of these
methods including SAM and SMA has become popular with the
advent of hyperspectral remote sensing (imaging spectroscopy).
SAM determines the degree of similarity between two spectra by
treating the spectra as vectors in a space with dimensionality equal
to the number of bands[2]. Each vector has a certain length and
direction. The length of the vector represents brightness of the
target while the direction represents the spectral feature of the
target. Variations in illumination mainly affect changes in vector
length, while spectral variability between different spectra affects
the angle between their corresponding vectors[2]. SAM is
appropriate for species-level monitoring at the regional scale as it
is not sensitive to differences in illumination or albedo[3, 4]. This
is often a problem associated with airborne imagery. SMA on the
other hand is a subpixel classifier that determines the relative
abundance of materials that are depicted in multispectral or
hyperspectral imagery based
characteristics. The reflectance at each pixel of the image is
assumed to be a linear combination of the reflectance of each
material (or endmember) present within the pixel.

Spectral Angle Mapper (SAM) is a physically-based spectral
classification that uses an n-D angle to match pixels to reference
spectra. The conventional SAM classifier as applied in one of the
most common image processing softwares, ENVI, compares the
angle between the endmember spectrum vector and each spectral
vector in n-D space. Smaller angles represent closer matches to the
reference spectrum. Spectra further away than the specified
maximum angle threshold in radians are not classified. A single
reference spectrum assumed to be the unique identity of each target
is used for the classification. If region of interests (ROI) are
selected to represent the spectral set for each class, the mean
spectrum is computed for all the spectra in the ROI and used as the
reference endmember. This means that spectral variability within
each class, denoted as the intra-class variability, is not preserved.
on the materials' spectral

Page 2

Some studies have gone through a stepwise approach by first
classifying possible classes for each species before merging them
into one class[3, 5].

Within-species (intraspecies) structural and chemical variability
create important challenges in mapping savanna tree species. For
example, the savanna landscape across the Kruger National Park
(KNP), South Africa and, neighbouring private game reserves and
communal lands present varying
characteristics albeit with similar species composition. The high
intraspecies variability in the KNP region is caused by differences
in edaphic conditions (e.g. presence of gabbro and granitic
substrates within relatively short distances), rainfall, herbivory,
bushfire and human activities. These factors do influence the
vegetation structure (e.g. green leaf area index) and biochemical
composition (e.g. differing chlorophyll and nutrient concentrations
resulting from phenological differences).

The objective of this study was to examine probable factors
including intraspecies spectral variability and the spectral sample
size that could affect remote sensing of Savanna tree species across
a land-use gradient in the Kruger National park. We have explored
two classification approaches with spectral angle mapper: (i) using
a spectral library composed of one spectrum (endmember) per
species and (ii) a multiple endmember approach conventionally
called K-nearest neighbour classifier.

vegetation structural

Figure 1. Study area showing CAO image scenes


2. METHODS AND MATERIAL

2.1. Study area

The study area is located in the “lowveld” savanna biome in the
North-East of South Africa (Figure 1). Eight sites were chosen for
the study including two sites in the KNP, two sites in private game
reserves and four sites in communal lands. The species data used in
this study consist of tree species generally more than 2 m tall
identified and geo-registered using a Leica differential global
positioning system (GPS). Eighteen dominant species are
examined in the study. These include Acacia gerradii, Acacia
nigrescens, Combretum apiculatum, Combretum collinum,
Combretum hereroense, Combretum imberbe, Combretum zeyheri,
Dichrostachys cinerea, Euclea sp (E. divinurum and E. natalensis,
Gymnosporia sp (G. buxifolia and G. senegalensis), Lonchocarpus
capassa, Peltoforum africanum,
Pterocarpus rotundifolia, Sclerocarya birrea, Strchynos sp (S.
madagascariensis, S. usambarensis), Terminalia sericea and
Ziziphus mucronata.

2.2. Airborne imagery
Airborne hyperspectral data were acquired in May 2008 with the
Carnegie Airborne Observatory (CAO) system for eight land use
sites across the study area (Figure 1). The CAO hyperspectral
system acquired the spectral information in the 384-1054 nm range
at a spatial resolution of 1m. The data were atmospherically and
geometrically corrected by the CAO. The species point map was
overlaid on the imagery and the spectral profiles collected via
region of interest tool in the ENVI software.

Table 1. Classification accuracy for 18 savanna species using (A) mean
spectrum for each species training set as reference endmember and (B) all
training spectra for each species as reference endmembers. A.
gerradii(AG), A. nigrescens(AN), C. apiculatum(CA), C. collinum(CC), C.
hereroense(CH), C. imberbe(CI), C. zeyheri(CZ), D. cinerea(DC), Euclea
sp(ED), Gymnosporia sp(GY), L. capassa(LC), P. africanum(PA), P.
thonningii(PT), P. rotundifolia(PR), S. birrea(SB), Strychnos sp(SY), T.
sericea(TS) and Z. mucronata(ZM).
A
AG AN CACC CH CICZ DC ED GY
AG
7522040031
AN
0110110110
CA
1100113221
CC
113012 030432
CH
0010100000
CI
2310130110
CZ
00100315000
DC
0121110040
ED
4102011360
GY
19132855313
LC
1003225200
PA
1311000120
PT
000 180000 132
PR
1130000204
SB
0201101023
SY
3312100200
TS
0000010010
ZM
1401100121
sum
3337144612 283024 4327
Producer
accuracy(%) 21.2 2.7 0.0 26.1 8.3 10.7 50.0 0.0 14.0 48.1 4.2 0.0 61.1 73.7 1.2 8.3 0.0 6.7
Overall
accuracy(%)
16
B
AG AN CACC CH CICZ DC EDGY
AG
12412120101
AN
31810120022
CA
1040102010
CC
10026120351
CH
0101201100
CI
11132100020
CZ
00200020100
DC
5101030630
ED
32021204230
GY
02000001114
LC
0101020120
PA
1000011001
PT
4003000001
PR
0100000001
SB
0324122422
SY
1202210000
TS
0131014124
ZM
1000000100
sum
333714461228 30244327
Piliostigma thonningii,
LCPA PT PRSB SY TS ZM sum User accuracy(%)
44 38 18.4
01 9 11.1
0120 0.0
2365 18.5
006 16.7
0019 15.8
40 38 39.5
0014 0.0
1237 16.2
222 101 12.9
9027 3.7
0010 0.0
3064 17.2
2028 50.0
2116 6.3
3020 5.0
004 0.0
1117 5.9
5315 533
2
2
0
6
0
1
0
0
3
4
1
0
2
0
2
1
0
0
2
0
1
2
0
0
1
0
0
3
0
0
2
1
0
0
0
0
1
0
0
5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
3
0
0
0
1
0
0
0
6
8
4
6
1
0
0
1
0
0
1
0
5
0
1
0
1
0
0
1
0
1
12
1
8
20
1
0
12
0
1
2
2
3
86
11
0
0
1
0
0
18
14
0
0
0
0
19241212

LCPAPTPRSBSYTS ZM sum User accuracy(%)
10 32 37.5
10 35 51.4
009 44.4
02 52 50.0
209 22.2
1126 38.5
2037 54.1
3026 23.1
0143 53.5
0120 70.0
0025 32.0
007 42.9
0121 38.1
0216 75.0
4181 55.6
0017 17.6
39166 59.1
0511 45.5
5315 533
0
0
0
2
0
1
0
0
2
0
8
0
2
0
5
1
2
1
0
2
0
3
0
1
2
0
0
0
0
3
0
0
0
1
0
0
0
0
0
3
0
0
0
0
0
0
3
0
8
0
2
2
0
0
4
0
0
0
0
0
0
0
0
0
0
0
0
2
3
0
3
1
2
1
0
0
0
0
0
0
3
0
0
1
0
0
0
1
3
3
0
10
1
3
1
6
0
2
0
45
2
2
3
86
12
1
0
2
0
1924121812
Prod ucer
accu racy(% )
Overall
accu racy(% )
36.4 48.6 28.6 56.5 16.7 35.7 66.7 25.0 53.5 51.9 33.3 25.0 44.4 63.2 52.3 25.0 73.6 33.3
48

Page 3

2.3. Data analysis

The spectral data for each species were randomly split into the
training and test data in a 1:4 ratio. Subsequently, two types of
reference endmember spectra were used to evaluate the effect of
within-species spectral variability on the discrimination accuracy
of the tree species; (i) the mean spectrum of the training data set
for each species and (ii) all training spectra for each species in a
multiple-endmember approach (K-nearest neighbour classifier). In
the basic nearest neighbours classifier, each training spectrum is
used as a reference spectrum and the unknown (test) spectrum is
assigned to the class of the closest (i.e. spectrally most similar)
reference spectrum. SAM was calculated between spectral pairs for
the whole spectral range (384-1054 nm). A target spectrum was
classified as species x based on the minimum SAM criterion
between a reference spectrum of species x and the target spectrum.
The overall, users and producers classification accuracies were
subsequently, determined from the confusion matrices between the
observed and predicted data.


A
80
Producer accuracy(%)
Pearson r = -0.35
0
10
20
30
40
50
60
70
00.020.040.060.080.1 0.12
Intraspecies SAM

B
Pearson r = 0.18
0
10
20
30
40
50
60
70
80
00.020.040.06 0.080.10.12
Intraspecies SAM
Producer accuracy(%)

Figure 2. The effect of intraspecies dissimilarity based on
intraspectral spectral angle measure (SAM) on the producer’s
accuracy for A. classification involving the mean spectra as
reference endmembers and B. all training spectra per species as
reference endmembers

3. RESULTS

The K-nearest neighbour (k = 1) classification provided a higher
overall classification accuracy (48%) when compared to the
classification involving the mean spectra of the training samples as
reference endmembers (overall accuracy = 16%) (Table 1). Seven
of the eighteen species showed producer’s accuracies of ≥ 50% for
the K-nearest neighbour classifier compared to three species for the
classification involving the mean spectra as reference endmembers.
Also, a higher number of species (8) showed user’s accuracies of ≥
50% for the K-nearest neighbour classifier compared to the
conventional SAM classifier (3 species).

The intraspecies SAM showed a negative but non-significant
correlation (r = -0.35, p = 0.15) with the producer’s accuracy for
the classification involving the mean of the training spectra as
reference endmembers (Figure 2). A significant negative
correlation (r = -0.58, p = 0.01) was however observed in the
above relationship when the analysis was conducted without
including one outlier. A highly non-significant relationship (r =
0.18, p = 0.47) was observed between the producer’s accuracy and
intraspecies SAM for the classification involving K-nearest
neighbour classifier (i.e. the multiple endmember approach).

The sample size of the training data showed a higher influence on
the classification accuracy when compared to the intraspecies
spectral variability or dissimilarity (Figure 3). A significant
positive correlation (r = 0.58, p = 0.01) was observed between the
producer accuracy and the training data size for the classification
involving the K-nearest neighbour classifier. No such influence
was observed for the classification based on the mean spectra of
the training sets as reference endmembers.

A
Pearson r = -0.15
0
10
20
30
40
50
60
70
80
0510152025 3035
Training data size
Producer accuracy(%)
B
Pearson r = 0.58
0
10
20
30
40
50
60
70
80
05 10 15 20 2530 35
Training data size
Producer accuracy(%)

Figure 3. The effect of training data size on the producer’s
classification accuracy for A. classification involving the mean
spectra as reference endmembers and B. all training spectra per
species as training endmembers

Page 4

4. DISCUSSION AND CONCLUSIONS

This study highlights two important factors that affect
discrimination of savanna tree species, namely, intraspecies
spectral variability and the training data sample size. The latter
being an important factor for the K-nearest neighbour classifier.
K= 1 nearest neighbours classifier was adopted in this study, which
minimises computational cost, but the classification accuracy may
improve with higher K values, particularly for classifications
involving large training samples. The complexity of the Kruger
National Park system requires the use of large training samples for
each species. The increasing accuracy of classification with
increasing training data size could imply that the classification of
the species could be limited to the dominant species. The criteria to
determine the dominant species have to be established.

We recommend the utilisation of the multiple endmembers SAM
approach as opposed to the traditional SAM classifier involving
single spectrum endmember per species for mapping of Kruger
National Park species and emphasis that the training endmembers
should be truly representative of the different distributions in the
population. Higher K values for the nearest neighbour classifier
with SAM may be required but the value of K must be as low as
possible to minimise the computing space requirement to store the
complete set of training data and a high computational cost for the
evaluation of new targets. Lastly, the spectral data for this study
was limited to the visible-near infrared. The utilisation of the full
spectral range from the visible to the shortwave infrared needs to
be assessed.

5. ACKNOWLEDGMENTS

Our gratitude goes to the Council for Scientific and industrial
Research (CSIR), South Africa and to the Andrew W. Mellon
Foundation for providing the funding for this study and to the
South African National Parks Board (SANParks), the scientific
services of the Kruger National Park in particular. Note of thanks
also goes to Barend Erasmus and Jolene Fisher of Wits University
and Konrad Wessels of the Meraka Institute.

6. REFERENCES

[1] J. B. Adams, M. O. Smith, and P. E. Johnson, "Spectral
mixture modeling: a new analysis of rock and soil types
at the Viking Lander 1 site.," Journal of Geophysical
Research, vol. 98 B8, pp. 8098-8112, 1986.
[2] 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," Remote Sensing of Environment, vol.
44, pp. 145-163, 1993.
[3] E. L. Hestir, S. Khanna, M. E. Andrew, M. J. Santos, J.
H. Viers, J. A. Greenberg, S. S. Rajapakse, and S. L.
Ustin, "Identification of invasive vegetation using
hyperspectral remote sensing in the California Delta
ecosystem," Remote Sensing of Environment, vol. 112,
pp. 4034-4047, 2008.
Y. Du, C. I. Chang, H. Ren, C. C. Chang, and J. O.
Jensen, "New hyperspectral discrimination measure for
spectral characterisation," Opt. Eng., vol. 43, pp. 1777-
1786, 2004.
J. T. Mundt, N. F. Glenn, K. T. Weber, T. S. Prather, L.
W. Lass, and J. Pettingill, "Discrimination of hoary cress
and determination of its detection limits via
hyperspectral image processing and accuracy assessment
techniques," Remote Sensing of Environment, vol. 96,
pp. 509-517, 2005.
D. A. Roberts, M. Gardner, R. Church, S. Ustin, G.
Scheer, and R. O. Green, "Mapping chaparral in the
Santa Monica Mountains using multiple endmember
spectral mixture models. ," Remote Sensing of
Environment vol. 65, pp. 267-279, 1998.
[4]
[5]
[6]