Research Of Anomaly Detection Algorithms Of Hyperspectral Imagery Based On Source Data Optimized

Hyperspectral imagery can distinguish different ground objects which just have tiny spectral difference by virtue of high spectral resolution. Also it doesn't need any prior knowledge of the target spectral signature. Thus it is very practicable in real scenes. Nowadays it becomes a hotspot in the field of target detection, and attracts many scholars'attention. Based on analysing of structure and characteristics of hyperspectral imagery, and applying morden signal processing techniques, the dissertation does the following researches in order to solve the difficulties in anomaly detection, such as high dimensionality, nonlinear correlation between spectral bands, and the strong interfere of background information to anomaly detection in mixed pixels.Firstly, based on research on dimension reduction methods, a new anomaly detection algorithm for hyperspectral imagery based on selective section principal component analysis is proposed reckoning on the characteristic of anomaly targets'distribution. According to block correlation of bands, principal component analysis technique is applied to remove correlation between spectral bands, and then according to local average singularity which can weigh up possibility of anomaly targets'existence, suitable principal components are selected for the following nonlinear anomaly detection by KRX. The proposed algorithm can greatly reduce the data volume of hyperspectral imagery, and effectively reserve anomaly target information.Secondly, the theory of linear mixture model is analyzed and a new anomaly algorithm for hyperspectral imagery based on background error accumulation is proposed. This algorithm extracts major energy information to be regarded as background subspace by principal component analysis in every subset of bands, and then obtains background error data which can suppress background and make target stand out for the following anomaly detection, through projecting hyperspectral imagery into that subspace. It can not only suppress interfere of background information to anomaly detection, but also utilize nonlinear information of imagery effectively. Thus, it can obtain a better detection result.Finally, reckoning on the degeneration problem of the background kernel matrix due to the background data blurred by anomaly samples in Kernel RX, a new weighted KRX algorithm based on target orthogonal subspace projection is proposed. The algorithm begins with the estimation of background covariance matrix, and structures target subspace through applying endmember extraction technique, then projects every pixel into target orthogonal subspace to give every pixel a proper weight adaptedly. Thus it diminishes the influence of background estimation due to the existence of target information, makes kernel matrix representate the real background data, and then increases detection probability.