Semi-supervised Dimensionality Reduction Methods For Polarimetric SAR Classification

Polarimetric synthetic aperture radar(Pol SAR) have all-weather, all-day, multi-band and other unique imaging characteristics and can provide a large area of high resolution imaging. By Further study on Pol SAR image processing and interpretation, we can extract more valuable information, and apply it to environmental monitoring, the earth resource exploration and military system, etc.Pol SAR itself did not bring more information about the ground, but can highlight some of the features of imaging target. For Pol SAR data, various polarimetric signatures can be obtained by target decomposition techniques, which are of great help for characterizing the land-cover. However, a single feature or feature set is not sufficient and reliable for discriminating different land covers that have a similar scattering mechanism, or the same classes of land covers that have different scattering mechanisms. So we adopt the strategy of collecting polarimetric features as much as possible to construct a polarimetric manifold by target decomposition, which aims to improve the classification accuracy. But polarimetric features may introduce redundant information and not all features for the land-cover classification is useful. So we need to dig out the essential features of classification effectively from a host of original features and reduce the redundancy between the features, namely for feature reduction(also called dimensionality reduction). In this paper, we put forward several semi-supervised dimensionality reduction methods for Pol SAR image classification, mainly includes the following three:1. A semi-supervised dimensionality reduction method based on spatial information is proposed. This method is the improved version for the supervised method LDLE(Linear discriminative Laplacian eigenmaps). The spatial information of the Pol SAR image has also been fully exploited. Experimental results show that this manifold learning based semi-supervised method demonstrates higher classification accuracy. In addition, the computational complexity is reduced by the matrix decomposition computing.2. We proposed a semi-supervised dimensionality reduction method based on subspace learning for Pol SAR image classification. This method firstly makes full use of the nearest neighbors of each pixel and polarimetric features as the form of a second order tensor. Then it takes advantage of the sample label information and manifold learning to construct the graph. And retain the polarimetric manifold structure of unlabeled samples. Finally, it calculates the eigenvalues and eigenvectors by solving the generalized eigenvector problem. The method has high classification accuracy for Pol SAR land-cover classification.3. A semi-supervised deep learning method based on spatial information is proposed. Firstly, this method uses Turbopixel algorithm to obtain superpixel blocks from pseudo color image and jointly takes advantage of the intersection of the ultra pixel block and the spatial neighborhood of the pixel instead of the pixel, can effectively solve label information of different categories on the boundaries of the classes. And then we apply the deep learning to learn feature and classify polarimetric SAR raw data(covariance matrix). Finally parameters are optimized by the gradient descent method by minimizing the cost function. The method implicitly extract more abstract, more effective features from Pol SAR coherent matrix for Pol SAR image classification and avoid mechanically extraction of features.