140 
Figure 3. Comparison of SW and BISE together with raw NDVI (before compositing) results for 
Fallow (a). Degraded Fallow (b). Millet (c), and Tiger Bush (d) sites. The solid lines are the BISE, 
dash lines are the SW, and the dots are the NDVI. 
Comparison of Classifiers 
Different classifiers as given in equations 1-4 were used in SW compositing algorithm to investigate the possibility of 
improving composited data quality by using alternative vegetation indices that were shown to be insensitive to external 
influences. The results of these classifiers are depicted in Figure 4 when applied to SW algori thm The temporal profiles 
of different classifiers are similar in general, but some of them still appear noisy. When the ground was covered by little 
vegetation canopies (before DOY 210), the GEMI and the NDVI appeared to be noisy while the others varied little. 
This is because of the sensitivities of these two classifiers were sensitive to soil background variations at low vegetation 
coverage or bare soils. As the canopy grew, the GEMI appeared to be smoother in temporal variations than the NDVI 
and similar to MSAVI and SAVI (Figure 4). In general, the valleys of the GEMI are coincident with those of NDVI, 
SAVI, and MSAVI, but occasionally are the opposite. At DOY 210 for Fallow, Degraded Fallow, and Millet sites, the 
GEMI had a valley while the others had small a peak, indicating the major discrepancies between the non-linear index 
and the ratio-based indices. The same behavior was found at the Millet site around DOY 270. 
All classifiers showed a contrast between the growing season and the pre-growing season, indicating all 
classifiers were sensitive to vegetation status. For qualitative studies of vegetation in the semiarid region, where soils 
dominate the remote sensing spectral signatures, the MSAVI and SAVI appeared to be better as they were less sensitive 
to soils, and therefore, were smoother in temporal profiles. However, for quantitative studies, the noises occurred in the 
temporal profiles of all classifiers tested need to be further quantified since all classifiers appeared to be noisy, although 
differed in degree. 
Figure 4. 
GEMI (w 
CONCLUDING R 
The MVC composi 
(Max Ch4), averag* 
within each compr 
consequendy restric 
algorithms may alst 
The increase in the 
also increase the ris 
profiles than other 
frequency) anomali 
which may last mo 
The slide i 
It retained more va 
algorithm is that it i 
earth surface in a fi 
algorithm is depend 
used. A good class 
however, all of co: 
registradon may be 
on "pixel-to-pixel"