Manolakis(2003) thinks that target detection is to distinguish the target from the background object and judge the existence of the target in each pixel. In the field of hyperspectral remote sensing, many target detection algorithms have been developed in recent years. According to the algorithm model, it can be divided into original space model, subspace model and albino space model (Zhang Bing et al., 20 1 1). Robey et al. (1992) put forward an adaptive matched filter (AMF), which simulates the background with a multidimensional normal distribution model, but it can't express the change of the background well. Liu Xiang (2008) analyzed the Elliptic Contour Distribution (ECD), and thought that this model could sensitively predict the changes of signals with the environment. When the target and background spectrum are known, Harsanyi( 1993) proposed an orthogonal subspace projection (OSP) algorithm, which considered the background spectrum and the maximum residual signal under various noises. In addition, Harsanyi also proposed a constrained energy minimization (CEM) algorithm. According to the target spectrum, the algorithm amplifies the signal in a specific direction and narrows other background signals, thus realizing target detection, which is suitable for small target detection. However, it is difficult for CEM detector to separate the target endmember signal from the noise signal (Bo Du, 20 10). Ling Lina et al. (2007) firstly deducted the background information of the image by PCA technology, and then selected the endmember by IEA (Iterative Error Analysis) method, and substituted the endmember spectrum into CEM as the known spectrum, thus extracting the small target well. In order to expand the application of CEM in large target detection, Geng Xiurui (2005) also improved the original operator by designing the weighted autocorrelation matrix, and proposed the CEM operator of the weighted autocorrelation matrix (WCM-CEM). Reed et al. (1990) developed the anomaly detection operator RXD, whose algorithm relies on the assumption that abnormal objects often drift out of the data "cloud" hyperplane constructed by image data, that is, when there is a large variance in the projection of the image in the direction connecting the abnormal pixels with the image mean vector, the operator will fail (Zhang Bing et al., 201/kloc-0 The algorithm whitens the hyperspectral data (WP), making the data "cloud" spherical in the feature space, while the abnormal pixels are still outside the spherical cloud, thus solving the problem of RXD detection failure. In order to expand the application of CEM in large target detection, Geng Xiurui also improved the original operator by designing a weighted autocorrelation matrix, and proposed a weighted correlation matrix CEM (WCM-CEM). Lin He et al. (2006) studied the hyperspectral data background and noise suppression methods of orthogonal subspace and target subspace projection. Lu Wei et al. (2006) proposed an unsupervised feature projection method, which extracted small targets from the perspective of abnormal distribution with the help of real-coded genetic optimization projection pursuit method.

In recent years, the scientific and technological circles and industrial departments at home and abroad have conducted in-depth research on mineral resources detection and heavy metal pollution monitoring in mining areas from different aspects. In the aspect of hyperspectral remote sensing detection of mineral resources, the spectral characteristics of rocks and minerals are measured by imaging spectrometer, the research of identifying minerals and detecting environment is carried out, and the integrated map information is obtained, forming the technical process and method of hyperspectral rock and mineral identification and mapping, and making breakthroughs in rock and mineral identification, information extraction and thematic mapping (Boardman et al.,1994; Du Peijun et al., 2003; Cruise et al., 2006; Jason, 2006; Zhang Bing et al., 2008; Wang Runsheng et al, 2007, 20 10). In recent ten years, research papers and reports on monitoring, analysis and evaluation methods of heavy metal pollution in mines have gradually increased. For example, using hyperspectral data and mineral identification pedigree to effectively identify pollution types in copper mines (Gan Fuping, 2004); The spectral characteristics of soil around coal gangue hill polluted by copper and heavy metals in different degrees were analyzed in the laboratory (Advanced, 2005). Based on the measured spectral data of spectrometer and considering the spectral characteristics of pollutants, the information extraction of pollution caused by mine pollutants and waste ore, water pollution caused by metallurgical wastewater and its vegetation pollution, and heavy metal pollution caused by long-term mining activities were studied (Kemper et al., 2002; Strictly observe honor, etc., 2003; Zhong Chang Kai, 2004; Gan Fuping, 2004; Cui Longpeng et al., 2004; Jason, 2006; Choe et al., 2008; Ren et al., 2009; Rashid, 20 10). In addition, some scholars have done a lot of research on vegetation biochemical parameters, vegetation index, derivative spectrum, red edge displacement analysis, regression analysis, stress effect, disease monitoring, pesticide residue detection, heavy metal pollution, etc. (Mutanga et al., 2004; Liu et al., 2004; Chen et al., 2009; Singh et al., 2010; Liu et al, 2011); With the in-depth study of the characteristics of spectral changes of ground objects in different environments, practical remote sensing quantitative detection technology of mine ecological environment based on subtle changes of spectral changes of ground objects has also appeared (Ferrier,1999; Advanced, 2005; Choe et al., 2008; Ren Hongyan et al., 2008; Jin Qinghua et al., 2009; Bech et al, 20 12).

To sum up, most of the existing achievements are to process and analyze the pixel spectrum or the measured spectrum of the spectrometer by means of extracting the characteristic points and parameters of the spectral curve, spectral differential processing, spectral absorption characteristic acquisition, spectral index calculation, statistical analysis, mixed pixel decomposition, spectral matching and so on. However, modern mathematical theories such as support vector machine (SVM), wavelet packet transform (WPT), harmonic analysis (HA) and adaptive neural network (ANN) are lacking in depth transformation of spectral curves, so there are great shortcomings in noise separation, foreign bodies in the same spectrum and foreign bodies in the same spectrum processing, and trace (weak) information identification. Therefore, it is necessary to carry out the application research of hyperspectral remote sensing data conversion and processing, information extraction and analysis based on modern mathematical theory.