Research On Mineral Spectral Feature Blind Extration And Target Detection In Hyperspectral Image

Target detection is a binary classification problem, which is to classify the image pixels into target and background ones. With the abundant spectral information of materials in hyperspectral image, the interest material can be detected for several applications, such as military reconnaissance, rock and mineral identification, environmental monitoring and so on. Compared with the traditional mineral identification, detecting targets from hyperspectral images has spectral advantage especially in rock and mineral identification. The geological mapping of a large area is accurate and fast by hyperspectral target detection. Besides, it is possible to classify and identify materials in the areas inaccessible for human beings. There are several problems in hyerspectral target detection, such as high dimension; same spectral with different matters and same material with different spectral; materials sometimes exist in the sub-pixel and the spectral information presented weakly.In this paper, we discuss the effects of blind source extraction method based on mean square cross prediction error(MSCPE_BSE) in mineral detection. The discussion is based on the introduction of linear mixed model and the present mineral spectral analysis methods. The main research work of the paper is as follows:Firstly, the present development situations of hyperspectral target detection, mineral detection and blind signal processing technology are introduced systematically, such as Constrained Energy Minimization(CEM), Adaptive Coherence Estimator(ACE) et al. Especially analyze the theoretical basis and implementation process of MSCPE_BSE. After the simulation experiments verification, compared with the results of classical detection algorithms, MSCPE_BSE method has a good detection effect.Secondly, a sub pixel target detection method, MSCPE-BSE based on subspace is proposed. Compared with MSCPE-BSE, the detection value is obtained by the orthogonal projection operator instead of the correlation coefficient between the detection and the target spectrum. Meanwhile, the subspace ideology is recommended, which can determine the position of the target during the subspace moving, and further research is carried out in the subspace contained the target. The ideology can reduce the scope, decrease the amount of data, and improve the efficiency of detection. Through MATLAB simulation experiment, compared with the classic detection method, this method can detect the target more accurately, and show better detection performance.Finally, the proposed method and the classical detection method are applied to the mineraldetection. The detection result figures, false alarm rate, ROC(Receiver Operating Characteristic) curve are as the evaluation indexes to evaluate these methods’ performances qualitatively and quantitatively. The experimental results show that, MSCPE-BSE method based on subspace can detect the target minerals in the mixed data which contain high similarity minerals, and with fewer false alarms, the detection performance has improved.