Research On Methods Of SAR Images Land Cover Classification Based On Local Features

SAR image land cover classification is the key step of SAR image auto-interpretation.Classified images can be used for the task of mapping,natural risk analysis and military reconnaissance,which have played a crucial role in both military and civilian.However,by reason of the speckle noise,geometric distortion,and being sensitive to system parameters and incident angle,SAR image classification is in face of the great difficulties.According to this problem,this thesis mainly studies methods of SAR image classification based on the local feature.The main work of the thesis includes following two aspects.(1).To solve the problem that no single feature can completely describe the target,a new feature combining the spatial and structural information is proposed.The proposed feature obtains the spatial information from gray-level co-occurrence matrixes and captures the underlying structure of SAR image by the method of local pattern histogram,and then the above two features are combined on the basis of considering the correlation between them.Finally,a kernel-based support vector machine is used with the proposed feature vectors for the classification of SAR images.Experiments on real SAR images demonstrate the effectivity of the method.(2).For the different tasks of classification,the paper makes some research on methods of SAR image land cover classification based on segmentation,and then a new algorithm for SAR image segmentation based on superpixel is proposed.In the method,superpixels are adopted as the operation elements,and then a new similarity is designed based on the gray and edge features.Using the new similarity,the similar superpixels are merged following the standard bottom-up approach.Finally,segmentation is completed via fuzzy c-means.Experiments on synthetic and real SAR images indicate that the method produces a better performance in vision as well as classification accuracy.