Compositing multitemporal remote sensing data.

Abstract

In order to reduce the problems of clouds, atmospheric variations, view angle effects, and the soil background variations in the high temporal frequency AVHRR data, a compositing technique is usually employed. Current compositing techniques use a single pixel selection criterion of outputting the input pixel of maximum value NDVI. Problems, however, exist due to the use of the NDVI classifier and to the imperfection of the pixel selection criteria of the algorithm itself. The NDVI was found not to have the maximum value under an ideal observation condition, while the single pixel selection criterion favors the large off-nadir sensor view angles. Consequently, the composited data still consist of substantial noise. To further reduce the noise, several data sets were obtained to study these external factor effects on the NDVI classifier and other vegetation indices. On the basis of the studies of these external factors, a new classifier was developed to further reduce the soil noise. Then, a new set of pixel selection criteria was proposed for compositing. The new compositing algorithm with the new classifier was used to composite two AVHRR data sets. The alternative approach showed that the high frequency noises were greatly reduced, while more valuable data were retained. The proposed alternative compositing algorithm not only further reduced the external factor related noises, but also retained more valuable data. In this dissertation, studies of external factor effects on remote sensing data and derived vegetation indices are presented in the first four chapters. Then the development of the new classifier and the alternative compositing algorithm were described. Perspectives and limitations of the proposed algorithms are also discussed.