Research On PolSAR Image Classification

Polarimetric Synthetic Aperture Radar(PolSAR) is an advanced imaging radar system with multi-parameter and multi-channel features. Compared with SAR images, the PolSAR data provides more surface information and fully polarimetric information. From now, PolSAR has become one of the most widely used detecting methods in the remote sensing field. PolSAR data also shows a massive growth trend. Therefore, PolSAR image interpretation has become a hot point in the PolSAR research field and PolSAR image classification is an important part of it. It plays an important role in both civil and military fields. This paper sums up the theoretical basis of PolSAR and some traditional PolSAR classification methods. After that, we proposed a new method combines tensor and the sparse auto-encoder. The main contents include the following aspects:First, for each pixel in the image, we use a there order tensor to express the initial data in the PolSAR T matrix. It covers the shortcomings that some methods use a scattering vector to express the information that will lose some data and change the nature structure of the PolSAR data. Our method mostly retains the information and the structure that will lead a good feature extraction and will provide better classification accuracy.Second, for each pixel in the image, we calculate the similarity between the pixel and its neighborhood to add the neighborhood information into the feature vector. It covers the shortcomings that some methods have not taken neighborhood information into consideration that will break region harmony. Our method mostly retains region harmony and improves the reliability of the classification results.Third, in this paper, for each pixel in the image, we use the sparse auto-encoder to extract features from the feature vector. It covers shortcomings that some methods have not considered the relationship between elements in the feature vector which reduces the learning ability. Using our methods in this paper can increase the learning ability by making use of the initial data. And the classification accuracy will also be improving.