The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
runs of test samples are around 88% and Kappa values of 0. 87
and 0.84, respectively. The species recognition accuracies
produced by ANN are acceptable when considering the
spectral similarity among most of the 11 species (Figure 2).
With exactly the same inputs of 30 selected spectral variables
as for ANN for both 11 species and 5 oak species recognition,
it is apparent that all accuracy indices produced by LDA are
very close to those by ANN, including both OAA and Kappa
values. The recognition accuracies generated by ANN and
LDA are not statistically different (Z<0.24), indicating, for this
particular case, that a non-linear recognition method does not
outperform a linear method. It might be due to most selected
spectral variables following a normal distribution by their
corresponding spectral samples.
Algorithm
Overall accuracy (%)
Kappa value
11 species
5 oak species
11 species
5 oak species
ANN
87.82
87.49
0.8656
0.8428
LDA
86.80
86.31
0.8546
0.8280
Note: The overall accuracies produced by ANN and LDA are not significantly different
at 0.95 confidence level for identifying either 11 species or 5 oak species
Table 3. Summary of species identification accuracy using
ANN and LDA algorithms with 30 selected spectral variables.
5. DISCUSSION
In this analysis, among the 30 selected spectral variables
evaluated by ANOVA, most of the spectral variables are
directly related to leaf chemistry, especially water status and
chlorophyll content in leaves. For example, some selected
spectral variables relate to water absorption bands: WI, DEP-
975, AREA-975, and Ratio975 directly relate to the 970 nm
water absorption band; Ratio 1200, NDWI, E-ID, DEP-1200,
AREA-1200, and WID-1200 are directly correlated with the
1200 water absorption band; and DSWI relates to the 1750
water absorption band. Other spectral variables relate to
chlorophyll content: C-1D, A-ID, B-1D, R550, A-WP, LCI, B-
WP, NDVI, SR, C-WP, and R680 that may directly describe
the variation of leaf chlorophyll content. In general, the full
range (350 nm to 2500 nm) of spectral wavelength covered by
the ASD spectrometer is useful for differentiating species that
differ in their foliage content, water status, pigment content
and other biochemicals, including visible, near infrared (NIR)
and middle infrared (MIR) regions (Nagendra, 2001).
However, due to the heavy water absorption bands near 1.4 pm
and 1.9 pm, which always happen in in situ spectral
measurement, visible and NIR bands are generally more useful
than MIR bands. In this case, most of the spectral variables
with high frequency of spectral separabilty between any
paired-species have been constructed from some of the visible
and NIR bands.
Although the recognition accuracy (around 88%) derived from
this study is acceptable for in situ species differentiation much
work is needed before applying this method to remote-sensing
image data, including high spatial resolution data, e.g.,
IKONOS (Carleer and Wolff, 2004) and QuickBird (Wang et
al., 2004) or hyperspectral data, e.g., AVIRIS (Xiao et a., 2004)
and HyMap (Buddenbaum et al., 2005). In this study,
atmospheric effects on the in situ hyperspectral measurements
was minimal except for the two major water absorption bands
in the MIR region. However, for remote-sensing image data,
atmospheric effects have to be corrected or compressed before
conducting a species recognition analysis (Nagendra, 2001;
Clark et al., 2005). Atmospheric correction will enhance the
spectral separability between species with multi/hyperspectral
remote-sensing data. Even so, the preliminary results with the
in situ hyperspectral data imply that current remote-sensing
techniques are still difficult but possible to identify similar
species to such 11 broadleaf species with an acceptable
accuracy.
6. CONCLUSIONS
The ANOVA analysis results indicate that the extracted 30
spectral variables are effective for differentiating the 11
species. The 30 selected spectral variables include spectral
variables related to water absorption bands: WI, DEP-975,
AREA-975, and Ratio975 (directly related to the 970 nm water
absorption band); Ratio 1200, NDWI, E-ID, DEP-1200,
AREA-1200, and WID-1200 (directly correlated to the 1200
water absorption band); and DSWI (related to the 1750 water
absorption band). The remaining spectral variables relate to
chlorophyll content: C-1D, A-ID, B-1D, R550, A-WP, LCI, B-
WP, NDVI, SR, C-WP, and R680 (may directly describe the
variation of leaf chlorophyll content). Both classification
algorithms (ANN and LDA) produced acceptable accuracies
(OAA from 86.3 % to 87.8%, Kappa from 0.83 to 0.87). In
this study, ANN and LDA for classifying the 11 broadleaf
species have a similar performance and the difference of
species recognition accuracies between the two classification
algorithms is not statistically significant at 0.95 confidence
level. The preliminary results of identifying the 11 species
with the in situ hyperspectral data imply that current remote
sensing techniques, including high spatial and spectral sensors’
data, are still difficult but possible to identify similar species to
such 11 broadleaf species with an acceptable accuracy.
ACKNOWLEDGEMENTS
This work was supported by University of South Florida (USF)
under the New Researcher Grant (Grant #: 18300). USF
graduate students, Mr. John Kunzer and Mr. John Merrill, are
greatly appreciated for their helping collect spectral
measurements in the field. Thanks also to USGS George R.
Kish for his valuable comments on this paper at an early stage.
REFERENCES
Asner, G. R, Palace, M., Keller, M., Pereira, R., Silva, J. N. M.,
and Zweede, J. C., 2002. Estimating canopy structure in an
Amazon Forest from laser range finder and IKONOS satellite
observations. Biotropica, 34 (4), 483-492.
Barnes, J. D., Balaguer, L., Manrique, E., Elvira, S., and
Davison., A. W., 1992. A reappraisal of the use of DMSO for
the extraction and determination of chlorophylls a and b in
lichens and higher plants. Environ. Exp. Bot., 32, 85-100.
Biging, G. S., Congalton, R. G., and Murphy, E. C., 1991. A
comparison of photointerpretation and ground measurements
of forest structure. In Proceedings of the 57th Annual Meeting
of American Society for Photogrammetry and Remote Sensing,
March, 1991, Baltimore,Md., 6-15.
Brockhaus, J. A., and Khorram, S., 1992. A comparison of
SPOT and Landsat TM data for use in conducting inventories
of forest resources, International Journal of Remote Sensing,
13: 3035-3043.
259