PolSAR Scene Classification Via Low-Rank Tensor-Based Multi-View Subspace Representation

Polarimetric synthetic aperture radar(Pol SAR)is a multi-channel synthetic aperture radar with four channels of HH,HV,VH,and VH.It is used in environmental monitoring,earth resource exploration and military systems.It is formed by transmitting and receiving electromagnetic waves vertically and horizontally.The polarization change of the backscattered wave is changed by reflecting the target,and the polarization sensor can obtain more information to describe the surface structure,surface coverage,geometric structure,etc.In addition,the Pol SAR system also has the advantages of working all day long,not being affected by weather conditions,and effectively capturing the ground geometry and geographic structure.However,the scene classification of Pol SAR images is still a rare topic in remote sensing data processing due to the difficulty of data acquisition.How to classify scenes of Pol SAR images is also an urgent task,and more time need to be invested in future research.Pol SAR images are rich in useful information for target classification,detection and recognition.Most of these information is only used in a single or simple combination in the processing of Pol SAR images.Multi-view feature spaces can comprehensively describe more information from different feature spaces,it needs to be better fused.By fusing the information in the multi-view feature spaces,the target recognition and classification will be more effective.Although using multi-view features can obtain sufficient information,how to mine the inherent characteristics between the multi-view features and reduce the redundancy of the multi-dimensional space is also a question worth considering.This article considers that tensor representation can be introduced into the algorithm.How to use these tensors to fully excavate their inherent characteristics in the multi-modal feature space is of great significance to the classification of Pol SAR scenes.At the same time,linear dimensionality reduction is also very important for processing Pol SAR images.Linear dimensionality reduction can reduce information redundancy and improve the classification accuracy of Pol SAR images.Based on the above characteristics of Pol SAR images,this paper classifies Pol SAR images by combining the multi-view features of Pol SAR images.The main description are as follows:(1)A low-rank Tensor-based Multi-view Subspace Representation(LRT-MSR)method is used to solve the problem of Pol SAR scene clustering.Pol SAR data can be described in multi-modal feature space,such as Pol SAR coherence,covariance,scattering,matrix,or various polarization decompositions.Different pseudo-color images from multiple spaces provide enough visual information for a comprehensive representation.The LRT-MSR algorithm uses a low-rank tensor subspace clustering method to explore the information of multi-view pseudo-color images.A tensor,as a high-order matrix,is used to capture the correlation of the underlying multi-view data.In addition,this method models the cross information from different angles under the constraints of low-rank terms,and obtains a series of self-representation matrices from the redundant information.Finally,the spectral clustering method is used for the final classification.Two new Pol SAR image data sets covering Shanghai,China and Tokyo,Japan were established using ALOS-2 fully polarization data for the experiment.The experimental results on two Pol SAR image data sets show the effectiveness of this method.(2)The dimensionality reduction algorithm improved by a low-rank subspace algorithm based on low-rank subspace representation(LRR)is used to realize the scene classification task of Pol SAR images.There is a large amount of coherent speckle noise in Pol SAR images,which destroys the special structure or the local geometric structure.As a result,many existing linear dimensionality reduction methods,including the popular linear dimensionality reduction method based on manifold learning,cannot be used in classification tasks.Therefore,robustness to noise in linear dimensionality reduction of Pol SAR images is an important issue.In this article,we integrate the optimal low-rank representation and projection learning into one model to enhance the robustness of low-rank.A novel robust image feature extraction framework is used in Pol SAR image scene classification,which improves the effect of Pol SAR image scene classification.(3)Inspired by the multi-view subspace clustering algorithm,combined with the projection learning image feature extraction framework,a low-rank constrained multi-modal tensor representation(LR-MTR)method is proposed.Using this method,combined with the multi-view features of the Pol SAR data,the scene classification of the Pol SAR image is realized.The LR-MTR method integrates Pol SAR data in a multi-view representation.In order to preserve space and polarization information at the same time,the features from different representation spaces(Freeman,H/A/?,Pauli,etc.)from a scene are formed into a cube(that is,a third-order tensor).On this basis,the low-rank norm is used to constrain the tensor from the multi-modal space to simulate the cross information from the multi-modal space.By minimizing the difference between the cascaded feature data and the corresponding feature subspace,the projection matrix is calculated.It also reduces the redundancy of these multi-dimensional spaces and solves the out-of-sample extension problem of large-scale data sets.Comparing with the most advanced linear dimensionality reduction algorithm,this method obtains the best quantization performance and shows superiority in fusion of multi-modal Pol SAR features and image scene classification.