Detection Of Small Target For Hyperspectral Imagery

With the development of spectrometer technology, target detection research in hyperspectral imagery has attracted more and more attentions in recent years. Hyperspectral imagery can capture spatial and spectral information simultaneously, and it has showed extraordinary advantages in the area of target detection, especially it has very important values in the detection and identification of military target. Due to the large amounts of data sets and small targets, some conventional target detection methods have no good detection performances in some applications. It's necessary to pursue some novel and efficient target detection algorithms used in hyperspectral imagery. In these cases, following aspects are researched in this paper.Firstly, this paper analyzes the data characteristics of hyperspectral imagery and studies statistics and mixture models of spectral signal. Local Gaussian statistical model and linear mixture model are mainly researched. This paper also describes multiple dimensionality statistical signal detection and estimation theory, which are the fundamentals of target detection algorithms.Secondly, three hyperspectral imagery target detection algorithms based on local Gaussian statistical model are studied, their detection performance analysis and simulation results are also given in this paper. In the adaptive filter bank approach, a practical method of designing filter bank is presented. To solve the disadvantage of RX algorithm which need target matched template, progressive RX algorithm-RXD is studied. Small targets are detected using RXD and other algorithms which are based on RXD. This paper presents a robus detection algorithm-RRXD that is based on RXD. Simulation results and ROC curves show that the detection performance of RRXD algorithm is better than that of RXD algorithm.Then, GMRF algorithm which is based on Gaussian-Markov random field model is studied in this paper. GMRF binary and single hypothesis detectors detected small targets, and their detection performances are analyzed. In this part, a new idea of incorporating the method of calculating GMRF inverse covariance matrix with other detection algorithm is introduced. LPD algorithm and CEM algorithm which are based on linear mixture model are studied. The two detection algorithms detected small targets successfully. According to the idea of LPD, this paper presents an adaptive detection algorithm-ACEM which is based on CEM. Simulation results and ROC curves show that the detection performance of ACEM algorithm is much better than that of LPD algorithm.Finally, this paper verifies RXD algorithm and RRXD algorithm by the use of DSP iamge processing platform. Small targets in the hyperspectral imagery are detected successfully, and the detection results show that detection performance of RRXD is better than that of RXD. Calculation speed of the two detection algorithms is analyzed, and some available meatures to improve calculation speed of detection algorithms are presented.