Research On Target Detection In Hyperspectral Imagery

Recently, with the fast development of hyperspectral imaging, hyperspectral remote sensing technology has been widespread concerned. Target detection in hyperspectral imagery can use both spectral and spatial information which contained by hyperspectral imagery. It owns incomparable advantages over the traditional space remote sensing images in target detection. Hyperspectral target detection can be utilized in diverse applications such as mine detection, defense and intelligence. On the basis of analysis of the structure and characteristic of hyperspectral image data, this paper focus on the interference of abnormal samples to background statistics, the building deviation of complex background model and the impact of background signal on sub-pixel target detection. The main work includes:Firstly, we analyze the characteristics of the hyperspectral image data. Given a short introduce to the models of spectral variability. This section focuses on the design flow of detection algorithms based on statistical detection theory. The multiple dimensionality statistical signal detection and estimation theory are also researched. This paper also describes the evaluation methods of hyperspectral target detection algorithms.Secondly, in the problem of contamination of background parameters estimation caused by outliers, this paper proposes an ACE algorithm based on robust background parameters estimation. The maximum likelihood estimation methods is mostly used in estimate the background parameters, but it is sensitive to outliers. Using these parameters modeling the background may have a big deviation due to outliers. The robust Fast-MCD parameter estimation method can enhance the accuracy and stability of the constructed background model and reduce the sensitivity of the detection algorithm to outliers. With experimental results on real datasets, the superiority of the algorithm is demonstrated.Finally, in the condition of the inaccuracy of modeling the complex background, this paper proposes an ACE sub-pixel target detection algorithm based on local background models. In the statistical detection algorithms, the more accurate the background models describes the real background distribution, the more effective the detector will be at separating the sub-pixel target from the background. The ACE target detection algorithm applies a single multivariate Gaussian distribution modeling the background. When the background homogeneity is high, the performance may be satisfactory. But it can't describe the real background distribution when the background is complex. In order to enhance the accuracy of the background model, we model the background a multiple Gaussian clusters. Each model represents a cluster with the same surface features, which can decrease the background model's spectral variability. For the effect of background signal to the sub-pixel target detection, we can get the abundance of background by fully constrained least squares. Apply the abundance to the ACE algorithm can suppress the influence of the background signal. Experiments show that this method has a better detection performance in hyperspectral image sub-pixel target detection.