Dimension Reduction For PolSAR Image Classification

PolSAR by observing different transceiver combinations under different channel polarization of echo signal,can provide more abundant information for feature analysis.It is an important research direction in the field of radar remote sensing.Through the transformation and decomposition methods such as available polarization feature of complex and diverse.However,too many feature combinations will bring computation burden and classification performance degradation,so dimension reduction becomes one of the effective solutions to the "dimension disaster" problem.Based on the major project of high-resolution earth observation system,this paper focuses on the feature and studies the dimension reduction methods of effectively reducing the label usage and realizing the migration application of training samples and training models,and the research results are classified in the PolSAR image classification practical applications.The main work and contribution of this paper are:(1)For the problem of feature redundancy in the practical application of PolSAR and the difficulty of marking sample size to satisfy the practical application requirement,the semi-supervised local discriminant analysis algorithm(SLDA)is proposed,and a framework based on SLDA for feature dimension reduction and PolSAR image classification was built.In the feature dimension reduction and the object classification task of three-view SAR image,SLDA can quickly extract low-dimensional feature with “in-class compact,inter-class separation” under the condition of 1‰-2‰ labeled samples,and the classification accuracy is about 90%.It solves the application problem with a small number of training samples and a very small number of labels in the need for rapid and accurate classification.(2)For the problem of feature redundancy in PolSAR application and the difficulty of marking sample size to meet the practical application requirements,an unsupervised dimension reduction algorithm--stacked denoising autoencoder(SDAE)was introduced.The multi-layer nonlinear network of SDAE can abstract the original data to more easily express the characteristics of its own properties and classify them.The classification accuracy obtained from the real scene polarization data of AIRSAR and gaofen3 satellite is higher than the traditional machine learning method.(3)For the problem of the mapping matrix trained by the dimension reduction method does not have mobility,and the training samples and the training model are not reusable in SAR image classification,a migration dimension reduction and PolSAR image classification framework based on SDAE was established.The experimental result of many real PolSAR scenes shows that the mapping matrix abstracted by SDAE in the polarization data of Radarsat-2 satellite Flevoland scene has strong generalization ability,which can be applied to the dimension reduction task of cross-satellite source cross-scenario,and the classification result is good.Therefore,it can realize the reuse of training samples and training models and reduce the workload of manual labeling and the cost of model training.