基于可见光/近红外反射光谱的沉积型稀土矿La元素含量无损检测  

Non-destructive Detection of La Content in Sedimentary Rare Earth Ores Based on Visible-Near Infrared Spectroscopy

作　　者：曹发生 刘严松[2,3] 何政伟[1,2,3] 刘心怡[1,2,3] 龚大兴 孙传敏[2,3] CAO Fa-Sheng;LIU Yan-Song;HE Zheng-Wei;LIU Xin-Yi;GONG Da-Xing;SUN Chuan-Min(State Key Laboratory of Geohazard Prevention and Geoenvironment Protection,Chengdu University of Technology,Chengdu 610059,China;Ministry of Land and Resources Key Laboratory of Tectonic Controlled Mineralization and Oil Reservoir,Chengdu 610059,China;College of Earth Sciences,Chengdu University of Technology,Chengdu 610059,China;Institute of Multipurpose Utilization of Mineral Resources,Chinese Academy of Geological Sciences,Chengdu 610041,China)

机构地区：[1]地质灾害防治与地质环境保护国家重点实验室(成都理工大学),成都610059 [2]国土资源部构造成矿成藏重点实验室(成都理工大学),成都610059 [3]成都理工大学地球科学学院,成都610059 [4]中国地质科学院矿产综合利用研究所,成都610041

出　　处：《分析化学》2021年第2期292-300,共9页Chinese Journal of Analytical Chemistry

基　　金：国家自然科学基金项目(No.41602120);中国地质调查局能源资源基地综合地质调查项目(No.DD20189507);四川省国土资源厅科技基金项目(No.J-2016-7)资助。

摘　　要：以黔西北二叠系宣威组沉积型稀土矿中的La元素为研究对象,对其进行光谱学分析的基础上,应用机器学习中遗传算法优化的极限学习机(Genetic algorithm-extreme learning machine model,GA-ELM)建模方法对La元素进行了高光谱定量估算反演。通过多种光谱变换(Savitzky-Golay平滑、微分、倒数、连续统去除等)后,使用了光谱特征细节凸显的反射率二阶微分研究对比了变量相关性系数(Pearson相关系数)和变量重要性评价(Competitive adaptive reweighted sampling,CARS;Successive projections algorithm,SPA;Radom frog,RF)两种特征变量选取方法所建立的数学模型精度差异。结果表明,沉积型稀土矿样中La元素含量与光谱反射率是一种非线性关系,不同的光谱变换方法对La含量信息提取能力不同,每种光谱变换都对应特定的敏感波谱区间。通过变量重要性评价方法所建立的模型比变量相关性系数所建立矿样La含量反演模型精度高。变量重要性评价筛选方法提取的特征变量反演模型显示,算法筛选的12个特征波长所建立的CARS-GA-ELM模型效果最佳,R 2(Determination coefficients)、RMSE(Root mean square error)和MRE(Mean relative error)分别为0.99、11.36 mg/kg、11.87%,能较好地检测La元素的含量。研究为“点”上沉积型稀土La元素的可见光-近红外光谱定量快速的反演研究提供了新的测试方法,为其它稀土土壤元素的光谱检测提供了思路,同时也为区域面积性的高光谱稀土资源的定量反演评价提供了理论依据。This work studied on La element,which hosted in the Sedimentary rare earth deposits of the Permian Xuanwei Formation,Northwest Guizhou Province.Based on the study of rare earth soil spectroscopy,the hyperspectral quantitative estimation and inversion of La element were carried out under GA-ELM model.After a variety of spectral transformations(S-G smoothing,differential,reciprocal,continuum removal),the accuracy differences of the mathematical models established by two methods of selecting characteristic variables(Pearson correlation coefficient and variable importance assessment(CARS,SPA,RF))were studied and compared.The results showed that there was a non-linear relationship between the content of La in mineral samples and the spectral reflectance.Different spectral transformation methods had different ability to extract the information of La content in soil.Each spectral transformation corresponded to a specific sensitive spectral interval.The accuracy of the inversion model of soil La content based on the method of variable importance evaluation was higher than that based on the correlation coefficient of variables.Among them,the best model was the first-order reciprocal of absorbance.The value of prediction R 2 reached 0.89,but the RMSE was 36.26 mg/kg,and the mean relative error reached 45.97%,showing a poor effect.The inversion model of characteristic variables extracted by the method of variable importance evaluation and screening showed that CARS-GA-ELM had the best effect,with R 2,RMSE and MRE reaching 0.99,11.36 mg/kg,11.87%,respectively,which could be well used to detect the content of La element.The research provided a new testing method theory for the fast and quantitative inversion of La element in soil on spot,showed a thought for the hyperspectral detection of other rare earth elements,and also provided a theoretical basis for the quantitative inversion and evaluation of regional hyperspectral rare earth resources.

关 键 词：沉积型稀土矿 特征变量选取 可见光-近红外光谱 定量反演 遗传算法优化的极限学习机 宣威组 

分 类 号：O657.3[理学—分析化学] P575[理学—化学]