Research On Robust Hyperspectral Anomaly Detection Method Based On Correntropy And Self-paced Learning

With the rapid development of remote sensing technology and sensor technology,the research on hyperspectral images is gradually deepened.Hyperspectral anomaly detection has also gradually become one of the popular researches in the field of remote sensing.However,hyperspectral images are inevitably affected by the acquisition equipment and environment during the acquisition process and contain mixed noise,which will seriously affect the accuracy of anomaly detection.How to improve the robustness of the hyperspectral anomaly detection method is the focus of this thesis.In this thesis,an in-depth study on how to improve the robustness of hyperspectral anomaly detection methods is conducted and two robust hyperspectral anomaly detection methods are proposed.First,an entropy rate and correntropy based collaborative representation detector is proposed.The method introduces the correntropy as a criterion for the reconstruction error between each test pixel and the corresponding background pixels to obtain the ability to suppress contamination and redundant bands in the spectral domain.Later,a background pixels set purification method based on the site entropy rate is proposed.The site entropy rate is used to measure the anomaly of each background pixel in the spatial domain,and the appropriate background pixel is reselected for collaborative representation according to the entropy rate.This method can effectively suppress the influence of noisy pixels in the spatial domain.Then,a self-paced collaborative representation with manifold weighting hyperspectral anomaly detector is proposed by this thesis.This method combines self-paced learning with collaborative representation to improve the ability of the model to suppress noise.This method divides each band of the hyperspectral image by the degree of noise contamination and adds the bands to the model in the order of low to high contamination.This method can prevent the model falling into a bad local minimum due to the interference of noise.In addition,to reduce the influence of potential anomalous pixels,a manifold learning reconstruction-based regularization matrix is proposed,which can extract the potential relationships between background pixels by manifold learning technique,and later adaptively assign weights to each background pixel by reconstruction errors.The two methods proposed in this thesis are evaluated on several hyperspectral datasets and compared with the current mainstream methods.The two methods proposed in this thesis outperform the compared methods in both qualitative and quantitative evaluation.This demonstrates the superior performance of the methods proposed in this thesis.