SAR Land Classification Technology Research Based On Polarimetric Information

Polarimetric SAR land classification is used widely, Which provides basis for targets verdict and recognition in the military area, such as for airdrome, tank and warship, and has a important application in crops inspection, environment detection and disaster prediction in the civil area. It has become the research hotspot of the remote sensing images classification area. Polarimetric SAR provides more abundant polarimetric information, however, which isn't enough in the land classification area's application and the classification algorithms have some shortages. Therefore, under this background, this dissertation gives a deep research in two aspects of the polarimetric information extraction and land classification algorithms.The main contributions of the dissertation are as follows:1.Analyze the development circumstance of the polarimetric SAR system and the technology of the polarimetric features extraction. Systematically summarize the historical decelopment and research circumstance in home and abroad of polarimetric SAR images classification technology from two aspects of supervised classification and unsupervised classification.2.In the aspects of polarimetric features extraction, extracts 6 polarimetric features from polarimetric scattering matrix, which don't change with rotation of the targets or variation of the polarimetric basis. Through the eigenvalue decomposition for the polarimetric coherency matrix, 6 features which reflect the polarimetric scattering mechanism of the targets are extracted. Besides, the deep analysis is given for the polarimetric covariance matrix and 13 features including polarimetric degree are extracted also. The whole feature set for the polarimetric SAR images classification is provided.3.For the problem of supervised classification algorithm of the polarimetric SAR images at present, this dissertation introduces the SVM theory, takes the average classification margin of SVM classifier as the evaluation criterion of the polarimetric features and uses the method of increasing l features or reducing r features to select the best feature subset. Compared with the Relief-F algorithm,the result is much better. Then, the polarimetric images are classified according to the best feature subset and SVM classifier. Experiment result shows that the classification accuracy is improved compared with the ML algorithm.4.Unsupervised classification algorithms based on Cloude decomposition for the polarimetric images are deeply studied. With the fixation problem of class numbers for the H/a/A-Wishart algorithm, this dissertation proposed a improved unsupervised classification algorithm,which uses classes'similarity to combinate the similarest classes. Experiment result shows that this algorithm accords with the factual land distribution better. For the Span/H/a/A-Wishart unsupervised algorithm,it needs a great deal of calculation amount in the process of classes clustering. So a detailed analysis for this problem is given and the Span/H/a/A-Wishart unsupervised classification algorithm based on polarimetric difference degree is proposed. Which substitutes the complex wishart machine for the polarimetric difference degree machine. Experiment result shows that this algorithm's calculation amount diminishes much more in the circumstance of the same classification accuracy.5.In order to offer experimental and applicational platform for the algorithms in this dissertation, a polarimetric SAR information processing module is programmed based on Interactive Data Language which integrates this dissertation's research results and some common polarimetric SAR information processing algorithms. Experiment result shows the validity of the algorithms in this dissertation. This module is practical and extensive.