Semi-Supervised Classification Of Polarimetric SAR Based On Graph

With the development of polarimetric SAR in theory and application, the airborne polarimetric SAR system has gradually developed into the spaceborne polarimetric SAR system. Compared with the single polarimetric SAR, the ability to solve practical problems have been significantly improved. On one hand, the polarimetric SAR can collect more information of the land covers. On the other hand, it is not restricted by single polarization.By using different polarimetric channels to obtain the target scattering properties, the observation becomes clearer and more accurate.After getting polarimetric SAR image data, the most important thing is the interpretation of it. Semi-supervised classification use unlabeled samples and a small number of labeled samples to improve the classification accuracy. With the extensive application of this method, the approaches based on graph model is also of concern, which is firstly using both labeled and unlabeled samples as vertexes to construct a graph. Each of edge in this graph is assigned a weight between each vertex. The weight value represents the degree of similarity between the two samples. By an optimization function, the category labels will spread to unlabeled samples by edges to finish the classification.Although the classification method based on graph has achieved a higher classification accuracy, there are two drawbacks: First, in the constructing process of these algorithms,ignoring the spatial information of the images has resulted in inaccurate classifications.The second is high time complexity in construction of the graphs. In this paper, to solve the above problems, we propose several semi-supervised classification methods based on graphs, and the main contents are as follows:(1) We propose a semi-supervised classification method based on the spatial-anchor graph to deal with the large polarimetric SAR data. First, referencing the key idea of anchor graph, according to polarization coherent similarity matrix of the sample, we used unsupervised Wishart clustering method firstly to group the data into multi-clusters. The cluster centers are served as anchors. Then each sample is represented by its nearest anchors to reconstruct graph. Combined with spatial information, the labels spread to unlabeled samples on the spatial-anchor graph, and the time complexity of the method has been largely reduced.(2) We propose a multi-view semi-supervised classification method on the basis of the single-view classification. This algorithm takes advantage of multiple groups of feature to construct several graphs. And then learning from others' strong points to offset one's weakness, we rerank the multiple graphs. The experimental results show that the proposed algorithm is effective.(3) We propose a fast semi-supervised classification based on anchor graph. We first divided area by using of superpixel segmentation, and then utilize a feature vector to represent each region to construct a single-view graph. After classification the category labels indicate all testing samples in the same region. By the proposed method, the running time is greatly reduced, while keeping a relative classification accuracy. The classification effeciency is improved.