基于光谱空间重构的非监督最邻近规则子空间的高光谱异常检测

The high dimension and huge data volume of hyperspectral remote sensing images and the complexity of surface feature lead to difficulty in distinguishing the anomaly pixel from the background. To solve these problems, an unsupervised nearest regularized subspace anomaly detection algorithm based on spectral space reconstruction is proposed. Firstly, in the process of band selection based on structure tensor, noise pixels are removed to obtain more effective bands. Then, the spectral space reconstruction is utilized to increase the absolute spectral distance between the background and the anomaly. Finally, to take full advantage of the spatial similarity information between background dictionaries, the spatial distance weight is introduced into the unsupervised nearest regularized subspace algorithm to improve the accuracy of linear representation.To validate the effectiveness of the proposed algorithm, experiments on four sets of real hyperspectral data are conducted, and the infulence of different parameters on the detection results is studied. Experimental results demonstrate that the proposed algorithm has a better detective performance than other anomaly detection algorithms.