Research On Hyperspectral Target Detection Based On Kernel Sparse And Spatial Constraint

Hyperspectral remote sensing provides detailed information of observed scene, also it is widely accepted that hyperspectral target detection has comparatively high research value. In military application, hyperspectral image can be utilized to identify important target. While in civil use, it plays a major part in environment monitoring, mineral resource location and precision agriculture. This dissertation takes advantage of the spectral and spatial information included in hyperspectral image using method of sparse representation and presents a novel target detection method with complete description of target object and high performance.First of all, starting with the basic theory and mathematical model of sparse representation, data requirement of norm relaxation and method of solving sparse equation are presented. With the analysis on sparsity and unification of image and spectrum, sparse representation model is applied to hyperspectral image. Based on the target detection method with sparse representation and experiment data, an elementary detection result is discussed.Secondly, with the consideration of the existence of non-linear feature and unification of image and spectrum in hyperspectral data, the optimization of target detection method is studied. Aiming to deal with the negative effect in detection result caused by non-linear feature, kernel method is introduced to the detection algorithm based on sparse representation. Thus the spectral information is deeply utilized to some extent. For the image property and spatial information contained in hyperspectral image, edge description and level set description are constructed to extract the spatial feature. Therefore, the spatial information is extracted and prepared for the application of target detection.Finally, considering the unity and integrity of spectral and spatial information in hyperspectral data, a joint sparse model based on spectral-spatial information is established in order to detect target object completely and accurately. With the help of the joint-information sparse model, amendment algorithms are proposed to revise the detection result from spectral information. Using AVIRIS hyperspectral data, two key parameters, sparsity and kernel value, are thoroughly discussed to investigate their effect to the detection result. Then incomplete data are created from original hyperspectral data. Classical detection methods, sparsity-based method and the proposed method are tested both on original data and incomplete data simultaneously. The outcome shows that the proposed method surpasses other target detection algorithms and indicates significant robustness under incomplete data compared with other methods.