Résumé
In GEOBIA, remote sensing experts benefit from a large spectrum of characteristics to interpret images (spectral information, texture, geometry, spatial relations, etc). However, the quality of a classification is not always increased by inserting a higher number of features. The experts are then used to define classification rules based on a laborious "trial-and-error" process. In this paper, we propose a methodology to automatically determine an optimal subset of features for discriminating features. This process assumes that a reference land cover map is available. The method consists in ranking the features according to their potential for discriminating two classes. This task was performed thanks to the Support Vector Machine-Ranking Feature Extraction (SVM-RFE) algorithm. Then, it consists in training and validating a classification algorithm (SVM), with an increasing number of features: first only the best-ranked feature is included in the classifier, then the two best-ranked features, etc., until all the N features are included. The objective is to analyze how the quality of the classification evolves according to the numbers of features used. The optimal subset of features is finally determined through the analysis of the Akaike information criterion. The methodology was tested on two classes of pastures in a study area located in the Amazon. Two features were considered as sufficient to discriminate both classes.