Neural Network Based Dunal Landform Mapping From Multispectral Images Using Texture Features

This paper presents a study towards machine generation of landform maps from optical remote sensing data. Our approach uses an offline trained multilayer perceptron (MLP) as a classifier, which is subsequently used to identify the landform classes in a satellite image. The paper emphasizes building a reasonably extensive database using multispectral images from which relevant texture information is computed. Gray level co-occurrence texture statistics, which form the feature vector representing the pattern, are used for training the MLP. Generalization results are assessed using the cross-validation mechanism. Performance of the algorithm is then extended to the problem of Aeolian (wind induced) landform mapping. Our results suggest that the textural method is promising for machine extraction of the landforms.