Fuzzy classification of spatially degraded Thematic Mapper data for the estimation of sub-pixel components

Abstract

Fuzzy classification procedures are hypothesized to provide more class membership information than hard methods as well as specific insights about sub-pixel components. This hypothesis is investigated in the present work using ancillary and spatially degraded Thematic Mapper (TM) data of a wide valley in central Italy. Hard and fuzzy modified maximum likelihood classifications (MLC) were applied to the data set. Fuzzy membership grades were found to be related to class cover proportions for each degraded pixel. Also, fuzzy probabilities were able to provide more precise estimates of class distribution, as measured by Kappa coefficients of agreement, with respect to hard attributions. Further analysis showed that this improvement was chiefly due to a better characterization of mixed, uncertainly attributable pixels. These results are of great interest for definition of the cases for which the fuzzy approach is appropriate.