Robust Kernel Joint Sparse Representation For Hyperspectral Image Classification

Hyperspectral remote sensing is an imaging technique that uses multiple and narrow electromagnetic wave bands to obtain relevant information of ground objects.Hyperspectral image data is an image cube composed of two-dimensional image space information and one-dimensional spectral information.Due to the high spectral resolution and rich spatial-spectral information,hyperspectral images have been widely used in many fields,such as agriculture,ecological environment,military affairs,and so on.Classification is a hot topic and also a key problem in hyperspectral image processing.Currently,spatial-spectral classifier that jointly uses the spatial texture information and spectral information is the research focus.Joint sparse representation(JSR)model is a representative spatial-spectral classification method.However,by minimizing the least squares approximation error under the linear and sparsity assumption,the original JSR model has the following problems:(1)the least-squares-based objective function is sensitive to noise and outliers;(2)the linear representation framework in the JSR model cannot describe the nonlinear characteristics of hyperspectral data.To solve these problem,this thesis focuses on robust and nonlinear kernel joint sparse representation models.First,we extent the original linear JSR model to kernel space and generate nonlinear kernel JSR(KJSR)model,which can effectively model the nonlinear spectral structural characteristics.Then,we change the least-squares metric to robust metrics,which are robust to noise and outliers.The research achievements help to enrich the theory of JSR,to improve the robustness,accuracy and practicability of hyperspectral image classification models,which have important academic value and wide application prospect.The main contents and contributions of this dissertation are as follows:(1)Considering the differences among neighboring pixels in the feature space,this thesis proposes a nearest regularized weighted kernel joint sparse representation(NRWKJSR)model.Different from the general weighted methods,the weights of neighboring pixels in the NRWKJSR are dynamically optimized rather than predefined.By simultaneously optimizing the weights of neighboring pixels and the sparse coefficients,the proposed NRWKJSR is more accurate and robust than the original KJSR.The optimized weights can highlight the similar neighboring pixels and weaken the effect of dissimilar noise or outlying pixels.(2)In order to eliminate the effect of noise,background and inhomogeneous pixels in spatial neighborhood,a maximum likelihood estimation based robust metric is proposed.Different from the traditional least-squares metric that is usually suitable for Gaussian noise,the maximum likelihood estimation based metric can handle different kinds of noise.Based on the metric,a maximum likelihood estimation based KJSR(MLEKJSR)is proposed.By alternate iterative optimization,the proposed MLEKJSR can be transferred to a weighted KJSR model.(3)To discriminate neighboring pixels,a self-paced learning strategy is embedded into the KJSR framework,which generates a self-paced KJSR(SPKJSR)model.The self-paced learning method can select neighboring pixels from easy to complex,which can effectively eliminate the effect of noise and outliers.In addition,we have evaluated the performance of proposed algorithms on the benchmark hyperspectral data sets.Experimental results demonstrate the effectiveness and robustness of the proposed algorithms.