Textural analysis methods for the classification of radar images

Presents a statistical approach to texture analysis. Texture is regarded as a two-dimensional random field defined by a suitable autoregressive model. Two methods are considered. The former employs a two-dimensional linear estimation technique: the grey level of a texture pixel is estimated from a weighted sum of grey levels of its neighbour pixels and the estimator that minimizes the mean-square error is used for texture characterization. The latter uses a simultaneous autoregressive (SAR) model, that characterizes spatial interactions of texture grey levels along fixed directions. Eight parameters corresponding with two different SAR models are extracted as textural features. These parameters capture texture characteristics in horizontal-vertical and diagonal-off diagonal directional pairs. These new techniques were applied to meteorological radar images, where precipitation and clutter regions correspond with two different and well distinct textures. Good results were provided by a minimum distance classification