PolSAR Image Classification Based On Deep Learning Method

Polarimetric Synthetic Aperture Radar(Pol SAR) is a kind of radar system of multi-channel imaging and regarded as one of the most advanced remote sensing systems in the field of remote sensing. Pol SAR data contains more information of terrain and target feature in its rich polarization information, owing to its multi-imaging mechanism, and the research in Pol SAR terrain classification has been the research focus in the field of radar image understanding and processing. In recent years, deep learning technology draws wide attention in machine learning field, and in the industry and academia have achieved amazing results. The paper mainly studies Pol SAR terrain classification based on deep learning method, and proposed an improved Pol SAR terrain classification method based on the statistical distribution characteristics of Pol SAR data. The main work is as follows:1. The deep belief network(DBN) in deep learning's method is investigated, and the basic structure and principle of DBN are analyzed. Emphatically the basic constitution unit Restricted Boltzmann Machine(RBM) is analyzed and Gaussian Restricted Boltzmann Machine(GRBM) is introduced. Based on GRBM, the classification of Pol SAR data is realized, and analyzes the classification performance of the algorithm through experiments in order to verify the effectiveness of the algorithm model in Pol SAR data terrain classification problem.2. Through the analysis of the principle of RBM, combined with the statistical distribution of Pol SAR data, which can describes Pol SAR data more properly. Wishart distribution based RBM is proposed, which we called WRBM for short. The experimental results show that WRBM can effectively learn the feature of Pol SAR data, and the classification results is better than GRBM based DBN and SVM, furthermore, WRBM based DBN get better classification accuracies.3. In the process of the DBN model, the fitting phenomenon exists in the training process, and a solution called Robust Training for DBN training is put forward. When the model is in the pre-training process, the neighborhood information is introduced, which is used to complete the RBM training through the reconstruction of the neighborhood information. In the fine-tuning stage of the model, the stability of the model is enhanced by adding the neighborhood information, which avoids the model from converging to local optima. The validity of Robust Training is verified by the experimental results.