Polarimetric SAR Image Terrain Classification Based On Deep RPCA Networks

Polarization SAR (Polarimetric Synthetic Aperture Radar, POLSAR) is a coherent multi-channel and all-weather radar imaging system. Compared with single-polarization SAR, fully polarimetric SAR can obtain more comprehensive feature information in favor of classification. The paper is combined with the deep learning methods which is based on polarization characteristics, the scattering characteristics and neighborhood information characteristics of the image data in order to further study classification methods of the polarization SAR image. The main work in this paper is as follows:1. The terrain classification method of polarimetric SAR image based on deep PCA(Principal Component Analysis) networks is proposed in the paper. Firstly, the data extracted from each pixel including the scattering characteristics, the polarization characteristics and the raw data elements of covariance, etc., is chosen as the input of deep PCA networks which can make fully use of the scattering mechanism and polarization characteristics of polarimetric SAR images. Then, we train the deep PCA networks and extract feature of the polarization SAR images from the bottom to the top, to ensure the validity and accuracy of the learning feature. Finally, calculate the precision. Experimental results show that the algorithm can further improve classification accuracy and the image quality of classification.2. Based on the scattering mechanisms of polarimetric SAR image, a terrain classification method of polarimetric SAR image based on deep RPCA (Robust Principal Component Analysis) networks is proposed in this paper. The normalized combination of the scattering characteristics, polarization characteristics and neighborhood information characteristics based on each pixel as the input to train the RPCA networks to which the ideal of low rank is applied to, followed by the SVM (Support Vector Machine) classifier for classification. The method takes full advantage of the scattering properties, the spatial characteristics and neighborhood characteristics of the polarimetric SAR data which is different from of other data, and further improve the robustness of the deep learning networks. Comparative experiments show that the proposed method has a higher objective evaluation and the classification performance.3. Considering the problem that the consistency of homogeneous regions keeps poor and too many training samples are existing, a terrain classification method of polarimetric SAR image based on the representation of superpixels, important sampling and deep RPCA networks is proposed. The training samples based on superpixel representation and important sampling are used to train the deep RPCA networks which changes the way for image classification from based on a single pixel into a superpixel. By the way, the classification is more precise and accurate, at the same time, the homogeneous regions keep more complete which also improves the classification accuracy of polarimetric SAR images and the image quality.