Sparse Deep Networks For Polarimetric SAR Terrain Classification

As more and more synthetic aperture radar systems appear, the polarimetric SAR data is obtained more and more abundantly. How to use these polarimetric SAR data to classify the ground scene has become a hot research topic. Deep learning, a new machine learning method, is developed in recent years. Because of its excellent learning and extension ability, it has been applied to many aspects.The present study investigates the sparse polarimetric encoding oriented to the terrain classification of polarimetric SAR and proposes an algorithm based on some practical problems to improve the existing algorithm.1. In order to solve the problems of the polarimetric SAR about data big and complex, a sparse polarimetric DBN oriented to the terrain classification of polarimetric SAR is proposed. The method realizes the terrain classification of the polarimetric SAR by incorporating rules of sparsity and combining the original polarimetric characteristics and the neighborhood polarimetric characteristics of the polarimetric SAR data on the basis of the existing DBN network. Such method makes full use of polarization information of the polarimetric SAR data, which makes the extracted features more representative, thus increases the accuracy and speed of classification.2. According to the speckle noise and huge amounts of data of the polarimetric SAR data, a sparse polarimetric autoencoding network oriented to the terrain classification of polarimetric SAR is proposed. This method realizes the terrain classification of the polarimetric SAR by adding a certain amount of noisy data and combining the original polarimetric charicteristics and the neighborhood polarimetric characteristics of the polarimetric SAR data on the basis of sparse self-encoding. Such method makes full use of the polarization information, which contributes to more robustness of the learned features and solves the problems of low accuracy of classification caused by the contamination of speckle noises in the original feature of the polarmetric SAR data, and thus increases the accuracy and speed of classification.3. In response to the unification of local class of the polarimetric SAR data, a terrainclassification of polarimetric SAR based on the superpixel segmentation and the sparse polarimetric autoencoding network is proposed. This method is to segment the Pauli PGB false-color image of polarimetric SAR into blocks with superpixel methods. Samples are selected from each block and then learned with the sparse polarimetric autoencoding network. The class of each block is decided with voting principles. Such method leads to a substantial improvement in the classification and region harmony.