of altitude z, slope s and shaded relief Sy44 of these pixels are 
170~290m, 12°~41° and 0.0~0.9 respectively. Owing to the 
topographic effect, the digital counts of the hardwood canopy 
varies from 34~40, 29~36 and 24~66 in XS1, XS2 and XS3 
bands respectively (Fig. 2). That such large variation of sensor 
response of a given canopy , especially in XS3 band, is the main 
reason to hamper the application of satellite image over 
mountainous terrain. The retrieved sea-level aerosol optical 
depth z,, (underline) are 0.46, 0.37, 0.44 and linear decay rate 
c, are 0.091, 0.071, 0.001 in XS1-XS3 bands respectively 
(Table 1). Mean hardwood reflectances 4 are 0.026, 0.023 and 
0.206 respectively in XSI-XS3 bands respectively. 
Sensitivity of derived surface reflectance to aerosol optical 
depth is also studied by artificially deviating the optimized z,, 
to +0.1. When z,,is under-estimated 0.1, 4 is over-estimated 
up to 0.049, 0.053 and 0.232 in XS1-XS3 bands respectively. 
On the contrary, 4 is under-estimated to 0.186 in XS3 band and 
even negative value ('fail' denoted in Table 1) in XS1 and XS2 
bands respectively. When tz, is deviated, surface reflectance 
at nearby shadow (S44; —0.0) is more sensitive to aersol optical 
depth than that at well illuminated as shown in Fig. 3. This is 
because diffuse irradiance determined from aerosol optical 
depth is dominated in nearby shadow terrain. The error of 
derived reflectance at shadow in rugged terrain can be up to 0.1 
in XSI-XS3 bands when Az, is 0.1, whereas the error is only 
about 0.01 in horizontal surface (Liu et al. 1996). Thus it 
could be concluded that more accurate aerosol optical depth 
should be needed in rugged terrain, if equal accuracy of surface 
reflectance is required. It is also noted that the optimized zs 
in XS bands don't follow Angstrom formula. z„oin XS3 band 
is greater than that in XS2 band. Probably it results from: (1) 
Lambertian surface assumption of the atmospheric correction 
model; (2) inaccurate modeling of adjacent slope reflected 
irradiance E, 
The surface reflectance is derived by correcting the atmospheric 
and topographic effects of the image. Table 2 shows the 
difference of mean reflectance in shaded and wel-illuminated 
(u,70.8) terrain before and after atmospheric correction. 
Acacia is chosen to verify the result of correction and testify the 
sensitivity of the algorithm to selected canopy. Typical 
spectral reflectances demonstrate the effect of atmospheric 
correction. In addition, difference of reflectances of pixels 
located in shaded and well-illuminated areas are greately 
reduced, especially in near-IR band. The reflectance 
difference of acacia in shade and well-illuminated area is -0.024 
in XS3 band, which is largely compared to the hardwood (- 
0.001) selected to retrieve the aerosol optical depth in the robust 
algorithm. Thus the algorithm is considered to be sensitive to 
the selected canopy. It is again probably due to the 
insufficiency of Lambertian surface assumption. 
By using the atmospheric correction model with retrieved 
aerosol optical depth as input, topographic effect is largely 
corrected in surface reflectance image (Fig. 1b) as compared to 
apparent reflectance image (Fig. la). To view the difference 
between with and without correction, the same enhancement 
function is applied. However, there are still some pixels 
under-corrected or over-corrected both at ridge or valley and 
intersection of bright-dark area. It is due to insufficiency of 
DTM spatial resolution or resampling of geometric correction in 
such a drastic change terrain. Landcover change may be one 
of the reasons as shown in the lower-left part of Fig. 1b. This 
area is over-corrected because its shaded relief (Fig. 2) is under- 
estimated. 
108 
Classification results bewteen with and without correction are 
also used to testify the algorithm. ISOCLASS clustering 
algorithm (IDIMS, 1992) is applied in classification. 
Uncorrected image is clustered to classes of urban, high- 
reflective land, mixed bare soil and urban, and two terrain 
related classes such as forest under high illumination, forest 
under low illumination (Fig. 4a). These spectral classes didn't 
correspond with real ground canopies very well, especially two 
terrain related forest classes due to topographic effect. 
However, forest, high-reflective land, bare soil, urban and grass 
are fairly well clustered if one uses the surface reflectance 
image to classify (Fig. 4b). The forest class includes 
hardwoods, acacia and tea, because their pairwise transformed 
divergences are all smaller than 1.6. The overall accuracy is 
91.7%. Kappa statistics is 0.87 (Congalton 1991). One can 
conclude that classification accuracy is improved if the satellite 
image is corrected by the atmospheric correction model. 
5. CONCLUSION 
Promising reduction of topographic effect of satellite image is 
achieved by using the proposed atmospheric correction model. 
Aerosol optical depth retrieved in the robust algorithm is shown 
to be optimized as deviated ones can produce large error in 
shadow areas. In comparing the sensitivity study of the 
previous study (Liu ef al. 1996), more accurate aerosol optical 
depth is needed to determine the surface reflectance in rugged 
terrain than in flat terrain. Classification accuracy is also 
improved by the corrected image. 
Although the results are rather encouraging, more studies 
should be undertaken: 
(1) verification of the aerosol optical depth by field 
measurement, and possible testify the model of adjacent slope 
reflected irradiance; 
(2) extension of the atmospheric correction model with non- 
Lambertian surface; 
(3) with DTM spatial resolution comparable to satellite image 
(Conese et al. 1993b), and possible with apparent DTM 
considering the tree-top of the terrain ( Liu 1995, Chen and Rau 
1993). This work is undertaken by the authors. 
REFERENCES 
Chen A.J. and JY. Chen, 1991, Using Lowtran6 and DEM to 
derive surface reflectance factor from SPOT HRV data. 
IGARSS'91.June3-6.Espoo, Finland,pp.651-654. 
Chen L.C. and J.Y. Rau, 1993, A unified solution for digital 
terrain model and orthoimage generation from SPOT 
stereopairs. /EEE Trans. on Geos. and Remote Sens.. vol.31, 
no.6, pp1243-1252. 
Civeo D.L., 1989, Topographic normalization of Landsat 
thematic mapper digital imagery, Photogramm. Eng. and 
Remote Sens., v55, n9, pp1303-1309. 
Colby J.D., 1991. Topographic normalization in rugged terrain, 
Photogramm. Eng. and Remote Sens.,v57, n5, pp531-537. 
Conese C., G. Maracchi, and F. Maselli, 1993a, Improvement in 
maximum likelihood classification performance on highly 
rugged terrain using principal component analysis, 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996 
C 
Hi 
H: 
H: 
He 
Iq 
Joi