A fuzzy spatio-temporal contextual classifier for remote sensing images

The proposed classification approach is based on a fuzzy fusion of three basic kinds of information: single-time posterior probability, spatial and temporal contexts. Such information derived from a multitemporal data set are exploited in order to improve the accuracy with respect to the single-time classification. Single-time class posterior probabilities are estimated by Multilayer Perceptron neural networks, which make the approach easily applicable to multisensor data sets. Both spatial and temporal contexts are derived from single-time classification maps provided by the neural networks. Expert knowledge about the possible transitions between classes at two different times is applied to the temporal context. The three kinds of information are then fuzzyfied in order to apply fuzzy reasoning rules to their fusion; fuzzy reasoning is based on "MAX" fuzzy operator and on the information about prior class probabilities. Finally, the class with the highest fuzzy output value is selected for each pixel to provide the final classification map. Experimental results on a multisensor (Landsat TM and ERS-1 SAR) and multitemporal data set consisting of two dates are presented. The performances of the fuzzy spatio-temporal classifier are compared with those obtained by a classifier based on Markov Random Fields (MRF). Results suggest that it represents an interesting alternative, which can be advantageous in particular from the viewpoint of simplicity.