Automated feature selection for MLP networks in SAR image classification

In object recognition using neural networks the correct selection of features is essential for achieving successful generalization of a net as well as satisfying time performance during the training and recognition phase. This paper shows the possibilities of automatically supporting this task in two steps. In the first step, a given feature set is examined with respect to its class separating capabilities. In the second step, the feature set is stripped of redundancies using the input pruning method introduced by Belue and Bauer (1995), which is applied to trained networks. Furthermore we show possibilities of extending these feature selection techniques by making use of context features, thus going beyond the scope of feature selection techniques known so far that only rank the features of the object to be classified. The application area we selected, is the pixel based object classification of SAR (synthetic aperture radar) images, where we use at present statistical features of the first and second order and some other texture describing features. The investigations are sponsored by Daimler Benz Aerospace, Dornier, who also placed the SAR image material at our disposal.