61
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
3.0
2 5
2.0
15
10
0 5
-Jt > . i I . i . L
4 6 8
Index number
1.0
vegetation
bare soil
water
impervious area
4 “ I 1 4 1 £
Index number
vegetation
water
shadow
impervious area
bare soil
Figure 4 Normalized spectral profile of endmembers extracted from pre-C corrected(left) and post-C corrected (right)ASTER
imagery in test site.
4. RESULTS AND DISCUSSION
4.1 Comparison analysis
In mountain area, the terrain undulation effect which could
cause shadow or occlusion is not ignored, shadows frequently
occur in airborne or space-bome imagery in terrain area with
steep slope when the sun elevation angle is low. A number of
technologies were developed to circumvent the terrain
illumination effect, here we selected the C correction method,
Fig2 shows the pre-corrected and post-corrected ASTER images
(bandsl32 for RGB)of test site. To visually compare the two
images same histogram stretch was applied to them and the
histogram of C corrected image was matched to the
pre-corrected one. As can be seen from Fig2, Fig2(a) had a lot
of shadows, the terrain undulation and stereo was apparent
across the scene(except the urban area), whereas Fig2(b) which
was C corrected had less shadow and the terrain undulation was
also inconspicuous. Fig2(b) was even more smooth than Fig2(b).
therefore it showed that the terrain undulation impact in study
site could be dramatically reduced by C method employed
above.
The spectra profile along a random line(Fig3) across undulate
area within test site showed that the spectrum variance of
objects which were similar in spectra value but locate in both
sunny side and opposite side due to the terrain undulation was
diminished(Fig3(b)). So the C correction could also effectively
reduce the phenomenon characterized by that same targets
display varying spectral radiance due to the topographic
undulation.
4.2 Regression and validation the C method
LSMM was applied to both pre-corrected and post-corrected
ASTER data of study site and two vegetation abundance
images were calculated. Unfortunately we did not collect the in
situ vegetation abundance data when the ASTER image was
acquired. But literatures showed that the vegetation abundance
and normalized difference vegetation index-NDVI or modified
vegetation index-MVI have close correlation(Carlson et
al,1997;Qi et al,2000;Zeng et al,2000 McDaniel and Haas, 1982),
the correlation coefficient could indicate the unmixing precision
at a certain extent. In addition, our main objective is to
investigate the impact on LSMM due to the topographic
undulation while not to specifically evaluate the precision for
unmixing the vegetation abundance, so the correlation
regression method is alternatively feasible for characterizing the
negative effect of terrain undulation in this study.
Normalized difference vegetation index and modified
vegetation index(MVI) or transformed vegetation index(TVI)
maps were derived from pre-corrected ASTER image as
follows:
NDVI=( R-NIR-Rred)/( RNIR-Rred) (7)
AM(TVI) =yj( RMR-Rredy( RMR-Rred)+0.5 (8)
Where R nir and R red are the spectral reflectances in ASTER
near-infrared(band 4) and red(band 3)bands;
To apply regression analysis 5,000 sample points were selected
randomly in four images including NDVI, MVI, vegetation
abundances from pre-C corrected and pos-C corrected ASTER
images. Sequentially the multi-regression analysis between
vegetation abundances and vegetation indexes(NDVI and MVI)
was employed and the results were displayed in Fig5.
Correlation coefficient (R)or square-R as a indicator index was
chosen to quantitate the unmixing precision indirectly. It should
be noted that a high R 2 value could indicate the good unmixing
precision and also the effectiveness of C correction. As is
shown in Fig5 the R 2 between vegetation abundances and
vegetation indexes followed by C correction is higher than that
of pre-corrected ones. The unmixing precision was improved
17.6% and 18.6% in R 2 value for NDVI and MVI by
minimizing or even removing terrain undulation effect using C
correction method. On one hand, it proved that, the terrain