Research On Anomaly Detection For Hyperspectral Imagery

Hyperspectral remote sensing is a new type of earth observation technology,hyperspectral imagery has the unique characteristic of acquiring spectral and spatial information simultaneously,it also has richer spectral and ground object imformation than multispectral imagery.Hyperspectral target detection is a research hotspot at this stage,it has a wide range of applications in agricultural production,enviromental protection,resource exploration,military reconnaissance,missile warning,etc.In practical applications,there are few target detection tasks that can obtain prior spectral information in advance,most target detection tasks are unsupervised anomaly detection.From a global perspective,an anomaly target is a small probability event in an image;from a local perspective,its spectral or spatial distribution mechanism is different from background samples.Anomaly detection algorithm need not obtain a priori spectral information,it is judged by the difference between the anomaly target and the background pixel.In this paper,anomaly detection in hyperspectral imagery is studied,based on the basic theory of hyperspectral target detection,and combining with the sparse representation model of signal analysis,the construction technology of background dictionary in anomaly detection task is studied in detail.The main research work of this paper is as follows:We introduce the classical target detection models of hyperspectral image,and analyze the basic principle and judgment criteria,then we compare some classical target detection algorithms which based on these detection models.From the qualitative and quantitative aspects,we introduces the indexes of evaluating the performance of hyperspectral target detection algorithm,namely ROC curve and AUC value.We introduce the anomaly detection algorithm which bases on sparse representation,and we analyze the problem of missing detection caused by the contamination of target atoms in the background dictionary learning process,then an improved anomaly detection algorithm based on sparse representation is proposed.The main measures of this algorithm are to construct the background dictionary based on RX operator,and to make a constraints on the sparse representation coefficient.Experiments show that this method can solve such problems and improve the accuracy of detection.We analyze advantages and disadvantages of the low-rank and sparse matrix decomposition algorithm and the kernel sparse representation algorithm,aiming at the optimization of background dictionary,a hyperspectral anomaly detection algorithm based on low-rank and sparse matrix decomposition and kernel sparse representation is proposed.The pure background matrix is obtained by the low-rank and sparse matrix decomposition algorithm,and the nonlinear features which can optimize the background dictionary are extracted by kernel method,then we use the reconstruction errors to achieve the detection purpose.Experiments show that this algorithm has better performance than several classical algorithms.