Posterior probability-based optimization of texture window size for image classification

Abstract

Texture provides spatial features complementary to spectral information in land cover classification of high spatial resolution imagery. In texture classification, window size is an important factor influencing classification accuracy, but selecting the optimal window size is difficult. In this paper, we propose an optimized window size texture classification method which can solve the window size selection problem. In order to validate the new method, we designed four classification experiments with different input features based on SPOT-5 imagery: (1) spectral features, (2) spectral features and single window size texture features, (3) spectral features and multiple window size texture features and (4) spectral features and optimized window size texture features based on posterior probabilities. Overall, the highest accuracy was obtained using the optimized window size texture classification, which does not require window size selection before classification. Furthermore, the results imply that optimized window size varies with land cover type.