Automatic division for pure/mixed pixels based on probabilities entropy and spatial heterogeneity

Mixed pixels have been one of the most formidable challenges in target recognition area of remote sensing, such as image registration and classification. Here we develop an automatic threshold selection approach based on belonging probabilities entropy and spatial heterogeneity. To perform the final automatic threshold selection, we employ kmeans to make a distinction between pure and mixed pixels. Experimental results in Tongzhou, Beijing using spot image confirmed the effectiveness of our method to convey pixel´s chaos. We further applied the division of pure pixels and mixed pixels to soft and hard classification and scale invariant feature transform (sift), some improvements are achieved.