Research Of Anomaly Detection Algorithms Of Hyperspectral Imagery

Hyperspectral imagery can provide a continuum of dozens or even hundreds of narrow bands, it enables us to detect the object of diagnostic spectral signature owing to its higher spectral resolution and richer images and spatial information than multispectral imagery.Anomaly detection algorithm does not require a priori spectral information by determining the pixels which are different from surrounding environment as anomaly targets, it has been a research hotspot in the field of hyperspectral target detection.In this thesis, hyperspectral anomaly target detection techniques and applications are mainly discussed. By systematically analysing the hyperspectral anomaly detection algorithm theory, the thesis focuses on problems such as extralarge data dimension, strong correlation and nonlinear between spectral bands, interference from background information, mixed pixels in hyperspectral anomaly detection. Specifically, the main research can be summarized as follows:(1) A hyperspectral anomaly detection algorithm based on minimum noise fraction (MNF) is proposed according to the characteristics of hyperspectral data. In this algorithm, the differences between the target and background are highlighted by reducing data dimension and separating noise by using MNF. The results of experiment on AVIRIS data show that the RX, KRX methods combined with the MNF possess good performances.(2) Aiming at the massive existence of mixed pixels in hyperspectral imagery, two spectral unmixing hyperspectral anomaly detection algorithm based on orthogonal bases approach (OBA) is proposed. Firstly we use the OBA method to extract the background endmembers, then suppress the interferece of the background information from the original hyperspectral imagery. By taking MNF transform and KRX method, The results of experiment on AVIRIS data shows it achieves a good detection result.(3) For hyperspectral imagery contains both abundant spectral information and extremely important spatial information, on the basis of the properties of kernel function, a hyperspectral anomaly detection algorithm based on the composite kernels is proposed. By combining the information of spectral dimension and the structural of spatial dimension, a variety of composite kernels can be formed. By extending this algorithm to KRX anomaly targets detection, the application areas ofKRX can be totally expanded. The results of experiment show that the kernel function combined with spatial information possesses stronger robustness.