The Polarimetric SAR Image Classification Based On Object Oriented Analysis

In recent years, polarimetric SAR classification is popular and it has great significance on the monitoring and identification of military targets in civilian areas as well as in the census on crops, land use and land cover, marine oil spills except in the military field. This article propose a new research idea which involved the optimal subset and the object-oriented analysis based on the Dalian City Radarsat-2 data and the coast of Newfoundland SIR-C data. The idea includes some components such as polarization decomposition, feature selection, OOA(Object-Oriented Analysis), Mahalanobis distance and SVM classifiers.This article first makes a introduction about the fully polarimetric SAR data. Then described its transformation from satellite microwave signals to computer images. Then, classify the various features of the SAR data into different categories. These are as follows:direct features which extracted from original data, features from target decomposition, and visual features of the image. For each type of feature extraction have introduced the use of reason and gives the algorithm. Again, for the features of three categories are divided into different subsets, a total portfolio of nine feature subsets were compared in order to obtain optimal subset, and reduce its redundancy. Finally, this article did object-oriented analysis to obtain ideal classification accuracy based on the optimal feature subset basis.Experimental results indicate that it does exist the optimal subset of features for different terrain and it can get better recognition performance and feature classification accuracy and it has its universality. Compared with formal classification methods, the introduced one can effectively reduce the speckel noise and improve the kappa. There are some disadvantages such as slowly speed, unsuitable for complex terrain for methods which available now. To solve these problems, this article employ a policy which starts with clustering segmentation and then merging with Lambda algorithm. The policy gets better performance. Because of the lack of training data,the policy needs to be testified one step further.