Abstract
The performances of a probabilistic contextual classifier and of an advanced per‐pixel classifier were tested in a sub‐tropical area of Western Africa using Landsat TM data. The results, evaluated both visually and statistically with respect to ground references, show the advantages and drawbacks of the two procedures in homogeneous and mixed, boundary zones. Since both procedures can produce probabilistic measures of class membership, a strategy is proposed for merging their positive features based on a weighted product of the output probabilities. Testing of the mixed approach on the same data set demonstrated its improved discrimination capabilities which are examined and discussed.