Research On Anomaly Detection Based On Kernel Method And Ensemble Learning In Hyperspectral Imagery

Hyperspectral remote sensing is an emerging science which is the first line in the remote sensing field. Hyperspectral image has the feature that it can provide spatial information and spectral information for the observed scene at the same time, so as the advantages of high spectral resolution and large amount of information, it has exhibit great potential value in many aspects of military and civilian application cause the capable of qualitative and quantitative analysis of target during detection. Hyperspectral remote sensing has the potential to offer multi-featured analysis. The subtle spectral information can be extracted by deeply mining the spectral feature space. The spectral features which reflect the target can help to increase the accuracy of target detection through comprehensive treatment of the spectral information. The multi-featured ensemble analysis is an important direction for the development of hyperspectral image processing, and the kernel method and ensemble learning theory developed recent years also have a good performance in the area of in Hyperspectral target detection and identification with its superior capabilities for nonlinear data processing, information extraction, small-sample learning, and high dimensional data high computing power. In this case, following aspects are researched in the thesis.First, the classic detection operators are studied from the statistical model and the small sample learning model respectively in the viewpoint of pattern recognition. The problems in the current anomaly detection algorithm are analyzed.Second, as the shortcomings of the current anomaly detection algorithm and the own requests of hyperspectral data, feature transform methods anomaly detection in hyperspectral image based on kernel method are proposed which can give a more effective expression for the anomalies in the feature space. With the combination of feature selection method based on local singularity metric and classical detection operator, the anomaly detection method are constructed base on anomalous component extraction and rare signal subspace estimation.Lastly, the ensemble learning theory is brought into the field of anomaly detection in hyperspectral images and a hyperspectral anomaly detection system based on ensemble learning is established. Through spectral division on the original data, sub-detector designation and decision-making, the multi-mode detection system have a more robust and generalized results than the traditional single-detector. Experiments show that the system is more capable to find the week anomalies which have small differences with the background.