Gaussian Maximum Likelihood and Contextual Classification Algorithms for Multicrop Classification Experiments Using Thematic Mapper and Multispectral Scanner Sensor Data

In this paper we present the results of a study of performance of a previously proposed classification technique on real remotesensor imagery. Testing has been achieved in the framework of an analysis of variance experiment designed to compare thematic mapper (TM) versus multispectral scanner (MSS) image data under the view-point of classification accuracy. The improvements of TM relative to MSS consist in (Fl) three additional spectral bands, (F2) increased radiometric resolution, and (F3) increased spatial resolution. The impacts of factors FI-F3, with or without context (factor F4), were evaluated by a four-factor analysis of the variance experiment, by repeated classification runs on 1) a TM data set, and 2) suitably degraded versions of the same set. Figures of increase/decrease in classification accuracy due to any combinations of the four factors have been computed, along with the corresponding levels of significance. Simultaneously acquired TM and MSS data sets have been used, together with photographic data acquired in coincidence with the satellite overpass (as control data for classification accuracy computation). The relaxation algorithms proposed in a previous paper have been used to assess the impact of the contextual factor.