Study Of Tensor Signal Processing-Based Classification Algorithms For Polarimetric SAR

As high-dimension data can be well represented by tensor due to its well preservation of the structure information in multi-dimensional data,tensor analysis has received much attention in remote sensing image processing for improved classification accuracy.Polarimetric synthetic aperture radar(PolSAR)image is one kind of typical high-dimensional image.For these reasons,this thesis studies tensor-based edge detection and classifier design methods for PolSAR image analysis.The contributions of this thesis are summarized as follows:Edge detection methods based on weighted structure tensor(ST)are proposed for PolSAR: Since traditional edge detection methods used for PolSAR are similar to edge detection methods for normal SAR,which usually produces unsatisfactory results,the average weighted ST is used in this thesis to detect PolSAR edges.It takes averagely weighting to combine structure tensors of each channel by assumping each channel provides the same edge information amount.Certainly,this may be not true in practice and thus,an eigenvalue-measured based weighted structure tensor is further proposed in this thesis for edge detection.It takes eigenvalue measure method to determine edge information amounts of each channel to formulate multi-dimension structure tensor.Experimental results show that,compared with traditional PolSAR edge detection methods,the proposed methods achieves better edge detection performance.Furthermore,this thesis proposes high-order representation of PolSAR sample and applies rank-1 support tensor machine(R1-STM)to PolSAR classification: The features for PolSAR classification usually originate in polarimetric decompositions.Due to the fact one kind decomposition tends to produce several features,there remains some structure information between the whole features from all kinds decompositions.Traditional PolSAR classification methods generally arrange the features as a vector,which will inevitably destroy structure information in feature set and thus be hard to obtain ideal classification results.To avoid this problem,this thesis researches PolSAR high-order representations and also applies R1-STM on PolSAR classification.Experimental results show that,the proposed high-order representation can indeed exploit structure information and consequently get better classification results.