The pixel selection criteria of current compositing algorithms are not specific enough to eliminate all of these
external factor-related noises as only a single criterion is employed. Besides, selecting one pixel with each compositing
period may ignore the real variations of NDVI due to vegetation changes such as anomalies. Consequently, pixel
selection based solely on the highest NDVI may not ensure high data quality and the MVC procedure may also result
in losses of valuable data.To improve composited data quality and restore data losses by MVC, current compositing
algori thms and the classifier need to be revised. The objective of this study is give a general review of existing
compo siting algorithms together with their classifiers and propose alternatives for compositing multitemporal remote
sensing data sets.
CURRENT COMPOSITING ALGORITHMS
Several algorithms exist for compositing multitemporal remote sensing data sets. Besides the MVC algorithm as
discussed above, other techniques have also been practiced by adding secondary pixel selection criteria such as minimum
c hanne l 1 (Min Chi), or ma ximum channel 4 (Max Ch4) of the AVHRR data (DTorio et al„ 1990; Goward et al„
1990). These techniques involve two-step pixel selections. The first stage of pixel selection retains a range (10 percent)
of maximum NDVI values within each composite period. Then, a second stage pixel selection is performed using a
secondary criterion.
The Min Chi secondary criterion selects the pixel with the minimum value in channel 1 from the AVHRR
pixels retained after the first stage of pixel selection. This technique is based on the reflectance characteristics of clouds.
Clouds have much higher reflectance values in ch anne l 1 than do other terrestrial surfaces. Selecting the pixels of
minimum channel 1 could further reduce the chances of choosing cloudy pixels. The remaining problem, however, is
the shadowing caused by clouds. The use of Min Chi as an additional criterion would choose those shaded pixels.
The Max Ch4 secondary criterion selects the pixel with the maximum value in channel 4 from the AVHRR
pixels retained after the first stage of pixel selection. This technique is based on the thermal properties of clouds. When
clouds or cloud-created shadows are present in a pixel, the thermal channel response of the AVHRR will be low.
Selecting the pixels exhibiting the maximum thermal channel within a range of NDVI would prevent the cloud and
shadow-affected pixels from being selected when better choices are available. Another approach is to set a threshold
on the temperature derived from channels 4 and 5 as the pixel selection criteria, which is not discussed further here.
Viovy et al. (1992) proposed a best index slope extraction (BISE) algorithm that uses vegetation growth pattern
as its secondary pixel selection criteria. The BISE examines each pixel and selects pixels according to whether the pixel
value matches the vegetation growth pattern. In a time series, the BISE searches forward and accepts the following day
pixel if the NDVI is larger than that of the previous one. A sudden drop in NDVI will be accepted, however, only if
there is no pixel, within a predefined period, that has an NDVI larger than 20 percent of the difference between previous
high and previous low values. If such a pixel exists, then the previous low value will be ignored; otherwise, the low
value will be selected. The rationale behind this is that the compositing classifier (NDVI) should follow the vegetation
pattern (steady growth followed by senescence). If any anomaly occurs, vegetation recovers slowly. Therefore, sudden
decreases in the NDVI classifier should be regarded as due to external effects unless there is a gradual increase in the
next few days. The BISE has advantages over the simple MVC because it retains more valuable data but does not ignore
sudden changes caused by anomalies. The disadvantages, however, are that this algorithm may select severely cloud-
contaminated data when clouds occur suddenly but disappear slowly.
Minimum View An\
surfaces. Selecting I
angles are preferable
could be minimized
keeps the maximun
From these remaini
variations among th
conditions. Conseqi
Because of the bidii
than the MVC.
Slide Window (SW,
pixel selections. Sta
one and will accept
both that pixel and
While searching for
there is a pixel wit!
there are no pixels \
value, within the sli
and the marked lov
values, the marked 1
from the date of th<
date.The advantage:
reduces the chance
than the length of t
Choice of Classify
Sensitivity analysis
studies. For compc
selected the follow:
1) Normal
2) Soil adj
3) Modifie
and
4) Global
where
ALTERNATIVES
Alternatives for Compositing
Noise in composited products was not only from using the NDVI, but also from the compositing algorithms used.
Therefore, alternative algorithms should be investigated to obtain reliable data products. The following ones are listed
as alternatives;
Average (AVG). In the first attempt, the maximum value, and down to 10 percent of the maxima, were averaged within
each compositing period. This algorithm uses a group value of vegetation index (VI) instead of using a single maximum
value. This is similar to a moving window average in that it uses mathematical averages, but it differs in that it only
averages those pixels whose values are within a certain range. These remaining pixels (10 percent) were presumably
not affected by clouds or poor atmosphere. Another underlying assumption is that the chances of large view angles are
equal to those of small view angles. The extremes induced by large view angles extremes would be averaged with other
pixels of small view angles. Consequently, the spikes due to view angle variation will be smoothed out, resulting in a
smoother temporal profile.
RESULTS AND (
Data Set Descripti
The data set used h
at Hapex Sahel stu<
and tiger bushes. P
November. The dat
NDVI profiles of ti
showed no significi
extremely high ref]
were due to line-dr
values for these pi
