Polarimetric SAR Image Classification Based On Spatial Information

Polarimetric Synthetic Aperture Radar has a very wide range of important applications and image classification is one of them. Many traditional classification methods are pixel-based methods. The performance of pixel-based methods are good, but it only uses the original information of image instead of spatial information, when used for high resolution image classification, the results may be not ideal.In this paper, we investigate the Polarimetric SAR image classification methods based on spatial information. They are accomplished by adding the space information into the use of semi-supervised classification method and the clustering algorithm in Polarimetric SAR image classification.1. We proposed a semi-supervised classification method based on multilayer features learning and spatial information of Polarimetric SAR image. In this method, there is only one original training sample in each class, it uses semi-supervised classification method of Self- Training, and the base classifier consists of framework of stacked sparse autoencoder and Softmax, which can learn multilayer features. Since of this method contains spatial information in selection of training samples, the performances are good and the accuracy are high.2. The previous method only contains the spatial information when choosing the training sample, but it does not use spatial information in the classification, so we proposed a Polarimetric SAR image classification method based on minimum spanning forest for its improvement. This method selects seed points according to the results of the classification of the previous method, the sum of the edges’ weights of minimum spanning forest is minimal among a set of all spanning forests rooted on seed points. In the same minimum spanning tree of pixels belongs to the same class label of root. The classification performances are better and the accuracy are higher.3. We proposed a Polarimetric SAR image classification method based on super pixels clustering by the method of fast search and find of density peaks, which can obtain the numbers of class adaptively. This method firstly uses Turbopixel algorithm to obtainsuper pixels for pseudo color image of pauli decomposition of polarimetric SAR image segmentation. Then calculate the Wishart distance between each two super pixels. The distances are as the input of method of clustering by fast search and find of density peaks to obtain the classification results and classes. After the clustering, let some super pixel class to zero by the post-processing. Then make all the pixels of each pixel block to be the same class. Finally, the zero class pixels are reclassified by supervised Wishart method. The method has high classification precision and is able to obtain the numbers of class adaptively.