Machine processing of LANDSAT data for soil survey: the Benue Valley savanna case study

Publisher Website
Google Scholar
Abstract
A ground-truth (GT) map, produced by fieldwork and the interpretation of large-scale black and white aerial photographs, was compared with digital/graphical output from LANDSAT data of the savanna around Makurdi in central Nigeria. The GT maps were digitized, grid-converted and then aggregated while the LANDSAT data were resampled for the purpose of rectification on the ERDAS 400 microcomputer at the Michigan State University. A visual comparison of the maps was done by overlaying the digital maps on the GT map using a colour monitor while the histogram listing provided an approximate quantitative comparison. The two algorithms of supervised classification (maximum likelihood and minimum distance) produced similar results but the third unsupervised classifier algorithm, cluster analysis, produced a far simpler map that is ideal as a reconnaissance soil/resource survey map. The major landforms were recognized by image processing but the reflectance-based classification resulted in misgroupings because of (i) the predominant influence of soil drainage regime on reflectance characteristics (thus well-drained soils tend to be grouped together just as were poorly drained soils, no matter what the intrinsic internal differences) and (ii) narrow units were masked by surrounding pixels and were therefore wrongly classified (a function of pixel size being dependent on LANDSAT's IFOV). The vegetation cover of the tropical savanna is a major problem in the digital classification of soils.