Hyperspectral Imagery Anomaly Detection Based On Collaborative Representation And Unsupervised Nearest Regularized Subspace

Hyperspectral remote sensing images contain a wealth of spectrum information of ground objects,which provide the conditions for the anomaly detection.How to automatically find a small number of anomalous targets in big data has become one of the hot research topics in the field of anomaly detection in recent years.To improve the detection accuracy of hyperspectral remote sensing imagery anomaly detection,the relationship between spatial distance and spectral variation in the local space,the comprehensive utilization of spatial information and spectral information,and the influence of outliers on the linear representation process are explored in this thesis.Some novelty strategies are introduced to improve the collaborative representation detection algorithm and unsupervised nearest regularized subspace algorithm,and the validation of the improved algorithms have been carried out on a synthetic anomalous dataset and two real image datasets The experimental results show that the improved detection algorithms have the stronger robustness and the higher detection accuracy comparing with the state-of-the-art.The main contents of this paper are as follows:(1)A double window sliding summation strategy is introduced in the hyperspectral images anomaly detection with linear representation model based on the reconstruction of background information.More spectifically,the background information around the test pixel has been fully utilized by expanding the single dual-window into multiple dual-windows.Moreover,the outliers have been restrained by inner window.Finally,local summation strategy is implemented to increase the utilization rate of spatial information in local background statistics,thus improving the accuracy of linear representation and the robustness.(2)According to the First Law of Geography,in hyperspectral remote sensing imagery,there are similarities among the spectral features of adjacent pixels,the closer the spatial distance between two pixels,the higher the spectral similarity.Considering of that correlation ship,the inverse distance weight is introduced in the current linear representation algorithm to make full use of the spatio-spectral similarity information,thus improving the accuracy of linear representation of the test pixel.(3)In the current hyperspectral imagery anomaly detection algorithms based on spatio-spectral information using background reconstruction,adding distance weight in the current algorithms could only reduce the influence of outlier on the linear representation process and cannot eliminated it.Therefore,a novel outlier removal strategy is carried out to improve the accuracy of representation in the unsupervised nearest regularized subspace algorithm.Specifically,summing the local background information obeys normal distribution,a certain confidence interval of the background pixels in the local window is found by statistical analysis,and then the removal strategy is implemented to eliminate the outliers in local windows to avoid the disturbing of outlier pixels and improve the accuracy of the test pixels with linear representation.