Research On Task Driven Hyperspectral Band Selection

A hyperspectral image is acquired by hundreds of spectral bands and contains rich information of many materials substances of interest,allowing effective identification of ground objects.Its great advantages make hyperspectral remote sensing technology applicable to many applications such as environmental monitoring,urban research and military reconnaissance.However,it is also faced with the issues of band redundancy,high spatial and temporal complexity and Hughes phenomenon,which poses a great challenge in how to process high-dimensional hyperspectral data for image analysis and interpretation.Therefore,for hyperspectral imaging,dimensionality reduction is necessary.Band selection(BS)is one of most important techniques because it not only contains the vital information of the data,but also retains the integrity of the spectral features,which is conducive to improving the analysis efficiency of hyperspectral data.As a result,BS has received considerable interest in hyperspectral imaging.Currently,most BS methods only focus on data characteristics and ignore the actual requirements of a particular task which results in the lack of spectral interpretation of the selected band subset,reducing the effectiveness of the method.Accordingly,this thesis is devoted to key issues encountered in classification,target detection and anomaly detection,and further develops the task-driven BS methods.Specifically,it integrates the process of hyperspectral BS by developing an adaptive method of determining the number of bands,band evaluation criteria and search strategies as well as an effective performance evaluation method.The main research focuses are as follows:(1)Class information-based hyperspectral BS.For image classification,a class information-based band number determination algorithm is proposed,which defines a new concept of using class features to evaluate the significance of each class according to information theory,and then designs two criteria,class self-information entropy and class entropy,to adaptively determine the number of bands required for single-class classification and multi-class classification.In addition,based on the constrained energy minimization(CEM)and linear constrained minimum variance(LCMV),the single-class signature-constrained BS and the multi-class signatures-constrained BS are proposed in conjunction with sequential search techniques,and the optimal band subset that contributes to classification is further obtained by a fusion strategy.The experimental results show that the proposed method can provide good guidance for the adaptive determination of the number of bands,and the selected band subset can actually achieve better classification.(2)Target-constrained interference-minimized hyperspectral BS.For target detection,by looking into the issues arising from target,interference and background interactions from the perspective of target detection and further exploring the relationship among bands,a targetconstrained interference-minimized filter(TCIMF)-based band selection(TCIMBS)is proposed.It selects an optimal set of bands in favor of target detection by evaluating the priority of single band or band subset.The experimental results show that the band subset selected by the proposed method can enhance the target detectability as well as the suppression of background interference to yield better detection performance.(3)Residual-driven hyperspectral BS.For anomaly detection,due to the lack of prior knowledge of anomalies,the anomaly detection-driven BS develops slowly.In this thesis,we put forward an interference-suppressed and cluster-optimized hyperspectral target extraction algorithm,which uses density peak clustering(DPC)to improve the noise immunity of the commonly used automatic target generation process and simplex growing algorithm to accurately extract target information and further provide a representative prior guidance for subsequent anomaly detection and BS.In particular,using an anomaly-background modelbased framework couple with orthogonal subspace projection(OSP),a residual-driven BS is proposed for unsupervised BS.The experimental results show that the proposed method solves the problem of missing anomalies to some extent,and can obtain a band subset with strong anomaly characterization ability and high stability.(4)A comprehensive three-dimensional receiver operating characteristic(3D ROC)curvebased analysis for evaluating classification performance.In order to make up the deficiency of traditional classification evaluation criteria,a classification performance analysis method based on 3D ROC curve is proposed.Firstly,a fractional class membership assignment mechanism is introduced to improve the hard decision of the traditional hyperspectral classification model,thus achieving better classification through continuous threshold segmentation.Secondly,the correspondence between the hypothesis testing model and the classification model is exploited to further extend the 3D ROC curve used for detection to classification.As a result,the evaluation issues caused by imbalanced classes and background for classification can be addressed.Finally,multiple evaluation criteria for detection and classification are integrated to formulate a performance evaluation score for assessment of classification performance.The experimental results show that the proposed method overcomes the limitations of a single evaluation criterion and provides a new approach to effective classification evaluation.