The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
257
position variables, and spectral absorption features were
extracted from the in situ hyperspectral measurements and
analyzed for classifying the 11 species.
Wavelength (nm)
Figure 2. Figure 2(a) shows a plot of unnormalized curves of
five oak species versus band wavelength for two observations
of each of the five oak species (Blue, Laur, Live, Sand and
Turk). Figure 2(b) shows the same curves of Figure 2(a) after
normalization. Figure 2(c) shows a plot of normalized curves
versus band wavelength for all the 11 species (Elm through
Turk).
3.3 ANOVA
To select a subset of spectral variables from the total 46 spectral
variables for running ANN and LAD for species recognition, an
one-way ANOVA analysis was performed. This was done
based on greater spectral separability between any two species
(paired-species) of the 11 species, using the SPSS statistical
package (www.spss.com, 2007). For any paired-species from
the 11 species, all spectral measurements for the paired-species
were used to conduct the ANOVA analysis across the 46
spectral variables (Table 2). Then based on the degree of
spectral separability of each spectral variable between the
paired-species, a statistical frequency was calculated at
probability levels p<0.01 and p<0.05 for each spectral variable.
For this analysis, a maximum frequency at either p<0.01 or
p<0.05 is 55 (because of n c 2 =55)-
3.4 Species Recognition Schemes
Two supervised classification schemes were employed for the
broadleaf tree classification: non-linear artificial neural network
(ANN) and linear discriminant analysis (LDA). In this analysis,
a feed-forward ANN algorithm was used for classifying the 11
species. An LDA classifier was also used to classify the 11 tree
species with inputs of the same subset of spectral variables as
for ANN to compare with the classified results by ANN. The
procedure DISCRIM in the SAS system (SAS Institute, 1991)
was used.
Two sets of samples were allocated - training and test samples,
from a total of 394 samples collected from 11 tree species. The
training samples were used for training ANN and LDA while
test samples are used to evaluate the tree species recognition
accuracies, generated with ANN and LDA. About 2/3 of the
samples were used for training and about 1/3 of the samples
were used as test samples. This procedure was repeated three
times (runs) (see Table 1) to obtain three different sets of test
samples (but training sets with a part overlaid between any two
sets). Finally, an overall accuracy (OAA) and Kappa index are
calculated from a confusion matrix produced with the test
samples using ANN and LDA.
4. RESULTS AND ANALYSIS
4.1 ANOVA
After the in situ hyperspectral data were preprocessed,
including smoothing and normalization, according to the
definitions for spectral variables listed in Table 2, the 46
spectral variables were extracted. An one-way ANOVA
analysis was first performed for all the extracted spectral
variables from which a subset of spectral variables was selected.
If the frequency threshold was set to greater than the half of
maximum possible frequency of 55, a total of 30 spectral
variables were selected.
Among the 30 selected spectral variables, all 10 Vis are
included, which imply that those Vis make a substantial
contribution to separating most of the 11 tree species. The 30
selected spectral variables can be further classified into two
groups. The first group of spectral variables mainly describes
the variation of foliage water content among the difference
species and its spectral variables consist of WI, DSWI,
Ratio 1200, E-ID, NDWI, DEP-975, AREA-975, H-WP, PRI,
DEP-1200, AREA-1200, Ratio975, and WID-1200. The