Fuzzy image classification for continental-scale multitemporal NDVI series images using invariant pixels and an image stratification method

The classification of multitemporal image data covering a large area is a challenging task because of scarce ground-truth data and the phenological variation of land cover in a study area. This research investigated an invariant pixel approach with an image stratification method. Using invariant pixels, the fuzzy image classification technique could be applied to every year with a satisfactory amount of ground-truth information with the AVHRR NDVI data covering Asia from 1982 to 1993. Invariant pixels were prepared by subtracting the growing-season average of the first three years from that of the last three years. A latitudinal image stratification method was investigated for minimizing the phenological difference of vegetation along the latitude. When land-cover information was transferred to other years by referencing invariant pixels, the classification accuracy of the other years showed only slight differences. The fuzzy classification results showed the decreases of forest and cropland areas, bur the increases of openlands such as deserts and rangelands.