The application of artificial neural networks and standard statistical methods to SAR image classification

In order to fully utilise SAR techniques, it is important to employ classification schemes which can discriminate between surface cover types having closely related statistics. A hybrid method, consisting of statistical textural measures and radial basis function (RBF) neural networks, is proposed for this problem. Imagery obtained for areas of South American rain forest are employed for this study and standard statistical techniques are used as a benchmark for comparison. A supervised method for training the RBF neural network hidden layer and parameters (e.g. centres, width, etc.) is proposed, based on a minimum-classification-error criterion. This modified RBF network has been applied to the forest data and has been found to outperform standard statistical techniques and the conventional RBF with k-means (or other similar) training method for hidden layer parameters in these classification tasks