PolSAR Image Terrain Classification Based On Deep ICA Networks

As far as now, most of features for Pol SAR(Polarimetric Synthetic Aperture Radar) image terrain classification are obtained by shallow machine learning algorithms, and these features are relatively low-level. When these features used to Pol SAR image terrain classification, the classification accuracy is not satisfactory. So, how to extract advanced features of the Pol SAR image becomes the focus and hot spots of research scholars. Based on the thought of deep learning, a deep network model based on ICA(dependent Component Analysis) is proposed. This model is used to extracte senior and separable features which can represent the Pol SAR image and then the extracted features will be used to implement terrain classificaton. When used in more Pol SAR image terrain classification, the deep ICA network model can get higher classification accuracy compared with shallow machine learning algorithms. The main contributions are as follow:1. A method for Pol SAR image terrain classification is proposed in this thesis based on features ensemble and deep ICA network. This method is mainly used to solve the problem that the existing shallow machine learning algorithms have disadvantages that extracting poor features and getting a low classification accuracy. This method is based on the thought of deep learning and cascades the ICA into deep neural network model. Ensembleing scattering decomposition features, texture features, color features of Pol SAR as the input of the deep ICA network to learn deep features represent. Then, we can get higher classification accuracy compared with shallow learning algorithms to confirme the feasibility and efficiency of the proposed method.2. In order to solve the problem that extracting scattering decomposition features, texture features and color features of Pol SAR takes too much time, a method for Pol SAR image terrain classification is proposed in this section based on K-Means neighborhood information coding and deep ICA. In this method, the ensemble features are replaced with neighborhood information of each pixels to improve classification efficiency. Then, KMeans is used to code the neighborhood information to reduce redundant information contained in it. Lastly, the coded results are used as the input of deep ICA network to get higher classification accuracy.3. Considering that the method in 2 have disadvantages that the classification accuracy is lower and the consistency of regions are not so satisfactory, a method for Pol SAR image terrain classification is proposed in this section based on deep Fast ICA and superpixels segmentat. The proposed method firstly classifies the terrains of Pol SAR based on each single pixel, then it combines the classification results with the segmentation results of Pol SAR to get the final classification results. This method can not only get higher classification accuracy than pixel-based classification method and shallow learning algorithms, but also better consistency of classification results.