Contribution of Variogram and Feature Vector of Texture for the Classification of Big Size SAR Images

Classical and modern statistical methods offer a wide variety of approaches to the classification of data in general and classification of imagery in particular. None of these approaches explicitly use spatial information. Spatial covariance structures have been used for data prediction, but not directly for classification. This paper describes a classification method using spatial covariance information delivered from texture in imagery to directly classify images in supervised approach. An experimental variogram is measured for each training zone, and the series of threshold points are sampled around the experimental variogram with the fitted models. Each point is checked to fit each of the theoretical variograms and the theoretical variogram with the best fit is chosen. In this study two models are used: the exponential and fractal models. With these two models (we have four parameters) in which each pixel is characterized by a feature vector whose components are known as Range, Sill, Slope and Fractal Dimension are constants deduced from the fitted models. This method has been applied on SAR image of the Atlantic coast of Cameroon. The proposed approach gives at mean 94% of precision. The proposal method also helps to gain in time of processing.