Graph-based Polarimetric SAR Classification With Improved Wishart Distance

Polarimetric Synthetic Aperture Radar(POLSAR) is an advanced technique to obtain remote sensing information. POLSAR obtains the Polarimetric information of the target from measuring the scattering characteristics from four different polarization combinations. Comparing to the traditional single-polarization SAR, the full-polarization SAR can get more completed information about electromagnetic scattering of ground target. The classification of Polarimetric SAR image plays an important role in the analysis and interpretation of Polarimetric SAR image. The classification result can be regard as an intermediate step which offers help to the edge extraction, target detection and targetrecognition of Polarimetric SAR images. And the classification can also be the direct need and the final result of the user. The traditional Polarimetric SAR image is based on pixels, which means the classification unit is a pixel. While there exits speckle noisein the Polarimetric SAR image, which highly impacts the classification result if the classification is based on pixels. In this paper, we construct a full-connected graph which contains vertex and edges with weights, and over-segment the Polarimetric SAR image into small regions. Merge the regions with small sizes hierarchically. Then classify the Polarimetric SAR image based on the regions. This paper includes mainly the following three sections:1 A graph-based Polarimetric SAR image segmentation is proposed in this paper. Firstly, extracting the Polarimetric features of pixels and construct a weight graph based on the Wishart distance. Then segment the Polarimetric SAR data into small regions based on the graph. Finally merge the regions hierarchically according to the size level of regions.We make different merging strategies for different size of regions. The algorithm introduces a graph-based segmentation from nature images, improves the weight measuring method with the Polarimetric information and takes the space information of pixels into account during the merging process, simple to think and easy to understand.2 Put forward a supervised graph-based Polarimetric SAR image classification algorithm. This algorithm is based on the regions segmented by the method mentioned above.Calculate the distance between a region and each training data, and classify the region to the class with minimum distance. The classification method is based on the regions which decreases the influence of speckle noise and increases the region harmony and the classification accuracy.3 Proposed a graph-based classification with the binary tree. This classification method is based on the segmentation regions. Calculate the differences between two different regions and construct the binary trees. The class number is the final number of binary trees. This algorithm improved the Wishart distance used to calculate the dissimilarities. This method decreases the local convergence and increases the classification accuracy of classification result.