稀散元素矿产资源是国民经济中的关键性矿产资源, 元素含量的提取是矿产资源潜力评价、 靶区优选的基础。 现有稀散元素含量分析面临快速检测、 潜力评价的需求, 基于高光谱的稀散元素含量反演是解决此问题的一种途径。 因此, 采集西藏斯弄多-则学矿集区的铅锌矿石, 开展铅锌矿石稀散元素镉(Cd)含量的高光谱反演方法与反演模型研究。 选用ASD Field Spec 3地物光谱仪及配套软件进行光谱数据采集和预处理； 在光谱特征分析基础上, 开展一阶微分(FD)、 二阶微分(SD)、 倒数的对数(AT)、 倒数对数的一阶微分(AFD)、 倒数对数的二阶微分(ASD)光谱数据变换处理, 结合皮尔森相关性系数(r)筛选特征波段, 进行随机森林(RF)、 人工神经网络(ANN)、 支持向量机(SVM)模型构建与反演, 选用决定系数(R2)和均方根误差(RMSE)评价反演模型效果与预测精度。 结果表明: 样品反射率集中于40%~60%区间； 1 420、 1 920和2 200 nm处形成吸收峰； 特征波段覆盖可见光和近红外波段, 771~2 051 nm为特征波段的最优区间。 SD光谱变换的降维效果最好, 筛选出15个特征波段； 其次为ASD和AFD光谱变换, 分别筛选出8个和2个特征波段。 FD与AT光谱变换未筛选出特征波段。 SD筛选的特征波段用于反演, 镉元素含量预测效果最好的是SD-ANN模型(R2=0.884, RMSE=2.679), 其次是SD-SVM模型(R2=0.830>0.8, RMSE=1.382), SD-RF模型预测效果最差(R2=0.505<0.6)。 ASD筛选的特征波段用于反演, 镉元素含量预测最好的是ASD-SVM模型(R2=0.857, RMSE=2.198), 其次是ASD-ANN模型(R2=0.846, RMSE=2.625)。 对比分析, 镉元素含量的高光谱反演模型效果为: SD-ANN>ASD-SVM>ASD-ANN>SD-SVM>ASD-RF>SD-RF。 该研究总结了铅锌矿石稀散元素镉的高光谱特征, 建立了镉元素含量的高光谱反演方法及模型, 为镉等稀散元素含量的高光谱反演、 无损检测、 快速分析提供了参考, 为高海拔勘探区稀散元素矿产资源的潜力评价、 靶区优选提供了科学支持。

Rare-dispersed element mineral resources are the key mineral resource in the national economy. The information extraction of content is the basis for potential evaluation and target optimization of mineral resources, but the existing analysis technology of rare-dispersed elements cannot meet the needs of rapid detection and potential evaluation. The analysis technology of rare-dispersed element based on Hyperspectral is a way to solve this problem. Therefore, the Sinongduo-Zexue ore concentration area in Tibet is the study area, and the hyperspectral inversion method and inversion model about the content of rare-dispersed element cadmium (Cd) in lead zinc ore are studied. ASD FieldSpec 3 spectrometer and supporting software are used for spectral data acquisition and pretreatment. Based on spectral feature analysis, multi-type spectral transformations such as first derivative (FD), second derivative (SD), logarithm of the reciprocal (AT), first derivative of logarithm of the reciprocal (AFD), second derivative of logarithm of the reciprocal (ASD) are carried out, the characteristic bands selected by Pearson correlation coefficient (r) are used for the modeling and inversion of random forest (RF), artificial neural network (ANN), support vector machine (SVM), the effect and prediction accuracy of content inversion models are evaluated by the coefficient of determination (R2), and root mean square error (RMSE). The results show that the spectral reflectance of the sample is concentrated in the range of 40%~60%, and the absorption peaks are formed at 1 420, 1 920 and 2 200 nm. The characteristic bands cover the visible and near-infrared bands, and 771~2 051 nm is the optimal range of the characteristic band. SD is the best spectral data dimensionality reduction method, which has selected 15 characteristic bands. ASD has selected 8 characteristic bands, and AFD has selected 2 characteristic bands. FD and AT did not select the characteristic band. In the characteristic band inversion of SD selection, the SD-ANN model (R2=0.884, RMSE=2.679) has the best prediction effect of cadmium content, followed by the SD-SVM model (R2=0.830>0.8, RMSE=1.382), SD-RF model has the worst prediction effect (R2=0.505<0.6). In the characteristic band inversion of ASD selection, the best prediction of cadmium content is the ASD-SVM model (R2=0.857, RMSE=2.198), followed by the ASD-ANN model (R2=0.846, RMSE=2.625). The hyperspectral inversion effect of cadmium (Cd) content is: SD-ANN>ASD-SVM>ASD-ANN>SD-SVM>ASD-RF>SD-RF. The study summarizes the hyperspectral characteristics of cadmium, establishes the hyperspectral inversion method and model of cadmium content, provides a reference for hyperspectral inversion, nondestructive testing and rapid analysis of rare-dispersed elements such as cadmium, and provides theoretical support for the potential evaluation and target optimization of rare-dispersed element mineral resources in the high-altitude exploration area.