A residual-based approach to classification of remote sensing images

This paper presents a novel residual-based approach to classification of remote sensing images. The proposed approach aims at increasing the accuracy of classification methods explicitly (or implicitly) inspired to the Bayesian decision theory. In particular, an architecture composed of an ensemble of estimators is used in order to estimate the residual errors in the class conditional posterior probabilities estimated by the Bayesian classifier considered. In order to avoid overfitting of the training data, a technique based on the analysis of class conditional entropy measures of omission and commission errors is used for adaptively evaluating the number of estimators to be included in the ensemble. Experimental results obtained on two multisource and multisensor data sets (characterized by different complexities) confirm the effectiveness of the proposed system.