Study On Anomaly Detection Algorithm Of Hyperspectral Imagery

Hyperspectral imagery is obtained by hyperspectral spectrometer which images objects in hundreds of narrow bands. Its bands range from the visible to the infrared. So hyperspectral imagery is a data cube that is made up of lots of two-dimensional images and every pixel in hyperspectral imagery has its own spectrum curve. The characteristics of hyperspectral data makes the study on hyperspectral processing becomes a hotspot in research. Anomaly detection for hyperspectral data is a target detection problem that no priori information about interesting objects is needed. It has a wide range of applications in fine agriculture, environmental engineering, defense and other aspects. This thesis is mainly focused on the study of anomaly detection in hyperspectral imagery, which has proved to be necessary and meaningful.Firstly, the classical processing algorithms for hyperspectral data are studied. Two algorithms for feature extraction are introduced: The first one is named after principal component analysis and the other one is kernel PCA. The classical anomaly detection algorithms are studied on the basis of the theory of multiple signal detection, the names of which are the benchmark RX algorithm, kernel RX and locally adaptive kernel density estimation. Their performance is analyzed by experiments.Then, considering the fact that anomalous pixels in hyperspectral data have a weak correlation with their surrounding background, the anomaly detection algorithm based on KCCA and SVD is proposed. Canonical correlation analysis is a kind of multivariate statistical analysis methods used to measure the correlation between two data sets. However, the use of kernel function can utilize the nonlinear information more effectively in hyperspectral data and kernel canonical correlation analysis is applied, which is the version of canonical correlation analysis in high-dimensional space. When the correlation between target and background in hyperspectral is analyzed by KCCA, singular value decomposition and reconstruction is used to complete the task of anomaly extraction and background suppression. Experimental results show the effectiveness of the proposed algorithm.At last, considering the nonlinear characteristics of hyperspectral data, the anomaly detection algorithm based on anomalous component extraction is proposed. In this algorithm, KPCA is used to do the preliminary work. Kernel principal components are extracted here. Independent component analysis is used to obtain independent component from kernel principal components. Two major methods for finding the most anomalous independent component are proposed here, which are different in the output order of the anomalous components. The first one is based on feature selection and the other one is based on orthogonal subspace projection. KRX detector is used to get the final detection result. The dataset collected from the airborne spectrometer is used to evaluate the performance of the proposed algorithm and the ROC curves provide a visual quantitative comparison. The experimental results indicate that the proposed algorithm has a better detection performance than the classical RX algorithm.