Hypersepctral Classification And Detection Methods With Sparse And Low Rank Priors

Hyperspectral remote sensing is a new remote sensing technique combining the remote sensing image and spectrum.Its spectral resolution can reach nanoscale which leads it to be one of the major technological breakthroughs in earth observation since 1980s.Hyperspectral data has been widely applied in modern military,mineral exploration,precision agriculture and environmental monitoring,and other fields.Therefore,efficient processing and interpretation of hyperspectral data have important theoretical significance and practical application value.The main contents of hyperspectral data processing are image classification and target detection.The abundant spectral information provided by hyperspectral data allows people to study and recognize the surface objects.However,the vast amount of information and the special data structure put forward severe challenges in image processing,information analysis,classification and detection etc.This requires people to understand and reveal the physical characteristics from multiple aspects.This thesis aims to propose some spectral-spatial classification and target detection techniques,design the corresponding effective algorithms on the basis of summarizing the state-of-art works of hyperspectral classification and target detection.In particular,the sparse and low-rank prior knowledge is discovered through the in-depth analysis of the characteristics of hyperspectral data.The main contributions of this thesis are as follows:(1)For the hyperspectral image(HSI)classification problem,an adaptive spatial context based joint sparse representation HSI classification method is proposed.Except for the abundant spectral information,every pixel in a HSI also has its own spatial structure.To make full use of the spatial structure information,the high order steering kernel for HSI is used to describe the local spatial structure.Then the joint sparse representation model and the local spatial structure are combined to get the final classification result.Experimental results show that adaptive spatial context based joint sparse representation classification method can effectively describe the surface object’s spatial context and improve the accuracy of HSI classification.(2)By effectively discovering the global and local spatial information in HSI,a low-rank decomposition based spectral-spatial hyperspectral classification method is proposed.First,an existing graph-based image segmentation algorithm is used to divide the HSI into many homogeneous regions.The matrix formed by the pixels in the same homogeneous region(a pixel corresponds to a column of the matrix)has strong column-wise correlation,thus it can be decomposed into the sum of a low-rank matrix and a sparse matrix.The essential feature is contained in the low-rank matrix,so the columns of the low-rank matrix are treated as the input features of the probabilistic support vector machine(PSVM)classifier.Besides,to elaborate the classification result,the Markov Random Field(MRF)regularization is introduced in the PSVM.By this way,the global spatial information,local spatial information and spectral information are combined efficiently.Experimental results show that the proposed method outperforms many of the state-of-the-art methods.(3)For the HSI anomaly detection problem,a low-rank and sparse representation based HSI anomaly detection method is proposed.The basic idea is that the HSI can be decomposed into anomaly part and background part.For the anomaly part,since there are only a few anomalous pixels,the anomaly part is modeled as a column-wise sparse matrix and the l2,1 norm regularization is used.For the background part,due to that the spectra in a HSI are drawn from multiple subspaces,the low-rank representation(LRR)model is used to find the lowest-rank representation of all pixels jointly given a dictionary.Then,a sparsity criterion is designed to characterize the local spectral structure of the pixels.Besides,due to that the atoms in the background dictionary should cover all the background materials and cannot be the anomaly pixels,a novel dictionary construction method is proposed.An effective algorithm is designed for the proposed model.Experimental results on both simulated and real data sets show that the proposed method can efficiently detect the anomaly pixels.(4)For the hyperspectral video sequences(HVS)gas plume detection problem,a spatial-temporal Total Variation(TV)regularized target detection method is proposed.According to that the gas plum in HVS has continuity in both spatial and temporal directions,the spatial-temporal TV regularization is introduced in the robust principal component analysis(RPCA)model.To make full use of the spectral information,PCA is adopted to extract the main features.Then,a novel fosion method is designed to combine the main features’ detection results.The experimental results show that the proposed spatial-temporal TV regularized model can describe the structure of the gas plume and improve the detection result.