A Mixed Multiscale Hurst Parameter Classification Model of Remote Sensing Image
一种遥感图象的混合多尺度Hurst参数分类模型

A classification model for remote sensing imaging is presented. The Extended self similar model(ESS)is a general fractional Brownian motion(fBm) model. Its multiscale Hurst parameters have relations with roughness. At the meantime, it doesn't require the roughness to be scale|invariant as fractal dimensions do. The ESS is closer to the realities than fBm. Moreover, ESS gives multiscale parameters to provide more accurate interpretation of textures while the fBH gives only one. The multiscale Hurst parameters can discriminate a large number of natural textures and are suitable to be the features for texture classification. These features' dimension is lower compared with many other texture features, so that the computation intensity is less. Directed Hurst parameters describe the roughness at four orientations and multiscales. In this study they are mixed with the mean and standard deviation of gray level to be the feature vector. A new classification model of mixed multiscale Hurst parameters is constructed based on Bayes theorem. In this model we suppose that the conditional possibility distribution function of each feature is Gaussian, and the features are independent with each other. The a priori possibilities are decided by the highest rate of correct classification of the training set. For remote sensing texture classification, the performance of the new model is compared to other features, such as co occurrence matrix features and Kaplan's features, etc. These classification algorithms are all based on Bayes theorem and the assumption that the a priori possibilities of all the classes are equal. Our experiments show that higher rate of correct classification to SPOT image is obtained by this new model.