137 
ite all of these 
h compositing 
quently, pixel 
lay also result 
it compositing 
:w of existing 
nporal remote 
' algorithm as 
:h as m inimum 
reward et ai., 
ge (10 percent) 
ormed using a 
n the AVHRR 
sties of clouds. 
I the pixels of 
m, however, is 
laded pixels, 
n the AVHRR 
: clouds. When 
R. will be low. 
the cloud and 
set a threshold 
d further here, 
growth pattern 
aether the pixel 
; following day 
iwever, only if 
tween previous 
;rwise, the low 
r the vegetation 
erefore, sudden 
increase in the 
does not ignore 
severely cloud- 
gorithms used, 
ones are listed 
tveraged within 
ingle maximum 
s in that it only 
ere presumably 
view angles are 
aged with other 
t, resulting in a 
Minimum View Angle (Min View). The NDVI tends to favor off-nadir in the forward direction for natural terrestrial 
surfaces. Selecting the maximum NDVI, therefore, creates a bias toward large view angles. In practice, nadir view 
angles are preferable because of reduced atmospheric path and bidirectional effects. Consequently, the view angle effect 
could be minimized by selecting the pixels with view angles nearest to the nadir. To implement this goal, the algorithm 
keeps the maximum and down to 10 percent of the maxima in each composite period (same as in AVG algorithm). 
From these remaining pixels, the one with the smallest view angle will be chosen. Again, the assumption is that 
variations among the remaining pixels are due to the view angles rather than due to the clouds or the atmospheric 
conditions. Consequently, the view angle effects will be minimized, and cloud-affected pixels will not be selected. 
Because of the bidirectional effects, this algorithm would, in general, tend to result in a lower vegetation index profile 
than the MVC. 
Slide Window (SW). Instead of using either the BISE or the MVC criteria alone, this algorithm combines the two in 
pixel selections. Starting from the first date, the algorithm searches forward. It compares the next pixel with the previous 
one and will accept it if greater than the previous one. When a vegetation index suddenly decreases, the algorithm marks 
both that pixel and the previous one. It then continues to search forward for a maximum of n days (slide window). 
While searching forward, it compares each pixel with the previously marked high pixel value. It will stop searching if 
there is a pixel with a value higher than the previously marked high pixel. Then a new search starts from that date. If 
there are no pixels with a value higher than the previously marked high pixel, the algorithm will compare the maximum 
value, wi thin the slide window period, with the previously marked low pixel. If the difference between the maximum 
and the marked low value is greater than 20 percent of the difference between the previously marked high and low 
values, the marked low value pixel will be ignored and the maximum will be selected. A new search will then continue 
from the date of the ma ximum . Otherwise, the marked low pixel will be selected, and a new search started from that 
date.The advantages of this algorithm are that it (1) keeps more valuable data than the MVC, (2) detects anomalies, (3) 
reduces the chance of selecting cloudy pixels, as in BISE, and (4) selects the cleanest pixels if clouds persist for more 
than the length of the slide window period. 
Choice of Classifier 
Sensitivity analysis (Qi et al„ 1994a) indicated that the choice of vegetation indices is dependent upon the purpose of 
studies. For compositing satellite data at regional and global scales for arid and semiarid land surface studies, we 
selected the following vegetation indices: 
1) Normalized difference vegetation index (NDVI): 
NDVI = ( p N1R - p red ) / ( p NIR + p rcd ) ; (1) 
2) Soil adjusted vegetation index (SAVI) of Huete (1988): 
SAVI = ( p N[R - p red ) / ( p NIR + p red + L)(1+L); (2) 
3) Modified soil adjusted vegetation index (MSAVI) of Qi et al. (1994b): 
MSA VI = {[ 2 p NIR +1 - /[ (2 p NIR +1) 2 - 8 ( p NIR - p red ) ] } / 2 ; (3) 
and 
4) Global environmental monitoring index (GEMI) of Pinty and Verstraete (1992): 
GEMI = ti ( 1 - 0.25 p ) - ( p rrt - 0.125 ) / ( 1 - p ^ ), (4) 
where 
= [ 2 ( p NIR ■ p" mi ) + 1.5 p NIR + 0.5 p rtd )] / ( p N r IR + p rcd ). (5) 
RESULTS AND COMPARISONS 
Data Set Descriptions 
The data set used in this study consisted of daily AVHRR data acquired during 1992 of four different vegetation types 
at Hapex Sahel study site (Table 1) near Niamey Niger. The vegetation types included fallow, degraded fallow, mill et, 
and tiger bushes. All possible 1.1 km resolution AVHRR data were acquired since early April till the be ginnin g of 
November. The data were geometrically, but not atmospherically corrected. The temporal reflectances in red, NIR, and 
NDVI profiles of these four study sites are presented in Figure 1. Although differed in vegetation types, these four sites 
showed no significant difference in reflectance before compositing algorithms were applied. The cloud-induced noises, 
extremely high reflectances in both red and NIR spectral regions, can be easily identified and those zero reflectances 
were due to line-drops. These noises can be reduced by compositing since the resultant NDVI classifier had very low 
values for these pixels.