CRF-Based Simultaneous Segmentation and Classification of High-Resolution Satellite Images

Abstract

Scale selection and uncertainty of image segmentation is still an intractable problem which influences the image classification results directly. To solve this problem, we adopt a CRF (Conditional Random Field)-based method to do segmentation and classification simultaneously. In this method, using probabilistic graphical model, we construct a three-level potential function which includes the pixels, the objects, and the link among the pixels and the objects to model their relations. We transform it to an optimization problem and use the graph cut algorithm to get the optimal solution. This method can refine the segmentation while getting good classification result. We do some experiments on the GF-1 high spatial resolution satellite images. The experiment results show that it is an effective way to improve the classification accuracy, avoid the boring segmentation scale and parameters selection and will highly improve the efficiency of image interpretation.