The Research Of Anormaly Detection Approaches Of Hyperspectral Imagery

Hyperspectral has become an important reconnaissance tool for military purpose. An anomaly detector can enable one to detect targets whose signatures are spectrally distinct from their surroundings with no a priori knowledge, and so it becomes increasingly important in hyperspectral image analysis. Based on the analysis of characteristics of hyperspectral imagery, the methods of anomaly detection are studied systematically in this paper.According to the mechanism of hyperspectral imagery, the principle of atmospheric factors, sensor factors and mixture of spectrum that cause the variability of image spectrum are presented in the thesis. The statistic properties and SNR of band image are analyzed, which provides the theoretical foundation for the construction of the anomaly detector and the creation of simulation data.Anomaly detection algorithms are studied based on full-pixel model. The influence of posteriori information on anomaly detection is discussed. The detector derived from Generalized Likelihood Ratio Test (GLRT) is developed under interference model. According to the geometry properties of anomaly in high dimension space, an algorithm based on analysis of inertia moment has proposed to detect image anomaly. The algorithm can detect image anomalies with small difference from the background spectrum.According to the principle of linear mixed model, we analyze the physical meaning and geometric meaning of endmembers in terms of anomaly detection. The performances of extracting geometry vertices and extracting mean spectrum are compared, and an improved 11:A (Iterative Error Analysis ) algorithm is proposed. In the thesis, we demonstrate the relationship between linear mixed model and the intrinsic dimensionality of data, propose the new idea of calculating the number of endmembers using intrinsic dimensionality.The anomaly detection approaches based on mixture model are classified in the paper; and the adaptive matched subspace detector is developed. The approaches based on Information-Processed Matched-Filter are studied. To solve the problem brought by the orthogonality constraints between the eigenvectors, a Low Probability Detection algorithm based on feature fusion has proposed. The approaches based on Full-pixel model and linear mixed model are compared synthetically using the emulation data and real data.Based on the analysis of linear mixed model theory and the lower-dimensional geometric properties of real data, we propose the idea that hyperspectral data always forms ahyper-plane in high-dimensional space, thus, according to the geometry characteristics of anomaly, we propose an approach based on the distance between points and the hyper-plane. The algorithm can solve the problem of mistaken detection because of image anomaly variety, improves the practicability of algorithms based on linear mixed model.