基于深度学习的花岗伟晶岩型锂铍矿物识别研究  

Identification of lithium-beryllium granitic pegmatites based on deep learning

作　　者：蒋果[2,3,4,5] 周可法 王金林[1,2,3,4,5] 白泳 孙国庆 汪玮[1,2,3,4,5] JIANG Guo;ZHOU Kefa;WANG Jinlin;BAI Yong;SUN Guoqing;WANG Wei(Center for Space Application Engineering and Technology,Chinese Academy of Sciences,Beijing 100094,China;State Key Laboratory of Desert and Oasis Ecology,Xinjiang Institute of Ecology and Geography,Chinese Academy of Sciences,rümqi 830011,China;Xinjiang Key Laboratory of Mineral Resources and Digital Geology,rümqi 830011,China;Xinjiang Research Centre for Mineral Resources,Chinese Academy of Sciences,rümqi 830011,China;University of Chinese Academy of Sciences,Beijing 100049,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院,空间应用工程与技术中心,北京100049 [2]中国科学院,新疆生态与地理研究所,荒漠与绿洲生态国家重点实验室,新疆乌鲁木齐830011 [3]新疆矿产资源与数字地质重点实验室,新疆乌鲁木齐830011 [4]中国科学院新疆矿产资源研究中心,新疆乌鲁木齐830011 [5]中国科学院大学,北京100049 [6]中国科学院空天信息创新研究院,北京100190

出　　处：《地学前缘》2023年第5期185-196,共12页Earth Science Frontiers

基　　金：中国科学院重点领域部署项目(ZDRW-ZS-2020-4-2);国家科学技术部新疆第三次科学考察项目(2022xjkk1306);国家自然科学基金新疆联合基金项目(U2003107)。

摘　　要：虽然遥感技术在大宗型金属矿产资源勘查方面取得了非常卓越的成效,但将其应用于稀有金属矿物提取的成果较少,尤其是对硬岩型锂铍矿物识别,主要受光谱分辨率、含矿岩体与围岩物性差异小、锂铍矿物光谱区分差异小等因素限制。为此,本研究通过野外采集含锂铍矿物伟晶岩和围岩样品并测量其光谱数据,使用光谱增强技术凸显光谱特征,对比分析特征吸收参数相似度模型和深度神经网络模型对矿物识别精度的影响。结果表明:(1)结合包络线去除和混合高斯模型提取的光谱吸收特征参数更简洁且具有更强的地质内涵;(2)光谱增强技术可提高模型识别精度,对比原始光谱,基于对数一阶导数光谱构建的模型的总体精度提高了0.05;(3)从总体精度看,深度卷积神经网络(总体精度=0.78)比浅层网络模型(反向传播模型总体精度=0.55和极限学习机模型总体精度=0.73)能够取得更好的效果。因此,结合高光谱技术和深度学习能够有效快速地识别花岗伟晶岩型矿物,为航空-航天成像高光谱影像直接提取锂铍矿物提供科学依据。Although remote sensing technology is widely used in large-scale exploration of metallic mineral resources,its application in direct rare-metal identification is limited,especially in the identification of hard rock Li/Be-bearing minerals.The problem is mainly due to low spectral resolution,low spatial resolution due to high physical similarity between ore body and wallrock,and small spectral difference between Li/Be-bearing minerals.To address this issue we investigate mineral identification methods based on deep learning models.Samples of Li-Be pegmatites and wallrock are collected from several pegmatite deposits and relevant spectral data are obtained.Spectral enhancement techniques are used to highlight the characteristic spectral features,and the characteristic absorption band similarity model and deep neural network models are compared for mineral identification accuracy.Results show that(1)the extracted characteristic absorption bands using a combination of envelope removal and mixed Gaussian model are better defined and reveal more geological insight.(2)Appropriate spectral enhancement can improve the accuracy of spectral models.In the case studied,the overall accuracy of the spectral model increases by 0.05 based on the logarithmic-first-order derivative spectrum over the original spectrum.(3)In terms of overall model accuracy,deep convolutional neural networks(0.78)perform better than shallow neural networks(0.55 for backpropagation;0.73 for Extreme Learning Machines).Overall,the combination of hyperspectral imaging and deep convolutional neural network model can quickly and effectively identify pegmatite-hosted minerals,which offer a scientific basis for the direct identification of Li/Be-bearing minerals by satellite remote sensing.

关 键 词：锂、铍 光谱变换 MICA 深度学习 高光谱 

分 类 号：P575.4[天文地球—矿物学] P618.7[天文地球—地质学] P588.121