A GENERATIVE METHOD FOR SIMULTANEOUS CLASSIFICATION OF REMOTE SENSING TIME SERIES DATA USING AN ENSEMBLE OF DECISION TREE CLASSIFIERS

Time series of remote sensing data has become an essential input for land use and land cover (LULC) studies. The current availability of multi-temporal data sets, from different sources and types, demands new classification approaches to explore their full capacity. In this study, we propose a non-parametric version of the Compound Maximum a posteriori classifier, based on an ensemble of Decision Tree Classifiers. This classifier was designed to avoid the classification of inconsistent class sequences in time. It was tested in a study area located in Itaituba, Pará state, Brazil, by the classifications of five Landsat images. In our case study, more than 25% of time series would be classified as invalid transitions. The use of the proposed approach substitutes these results with the most probable consistent class trajectory. Improvements in individual accuracies, when compared to post-classification comparison, have also been observed