矿产预测中的成矿因子选择方法:以滇东南金矿预测为例  

Feature selection approach in mineral prospectivity analysis:a case study of gold deposits in Southeastern Yunnan,China

作　　者：俞乐[1,2] 柏坚[3,4] 张汉奎[1] 

机构地区：[1]浙江大学地球科学系,浙江杭州310027 [2]清华大学地球系统科学研究中心,北京100084 [3]中国地质大学地球科学与资源学院,北京100083 [4]云南省地矿局,云南昆明650011

出　　处：《浙江大学学报（理学版）》2011年第3期348-353,共6页Journal of Zhejiang University（Science Edition）

摘　　要：由于矿产地质信息的复杂性和不确定性,难以建立精确的数学模型来确定矿产资源的分布状况.非线性分析建模技术,如人工神经网络(Artificial Neural Network,ANN)、支持向量机(Support Vector Machine,SVM)等,给矿产预测工作提供了新的途径.这类方法在处理数据时可以避免数据分析和建模的困难,即不须理解各种成矿因子与矿床(点)之间的相互关系,只须选择已知的矿床(点)和非矿产(点),进行"黑箱"学习.虽然经过合理的训练,这类方法能够得到较高的预测精度,但由于其分类过程的非线性特性,难以获得容易理解的分类规则,提供成矿因子的知识.本文采用基于SVM的迭代特征消去(Recursive Feature Elimination,RFE)技术(SVM-RFE),即在SVM模型的训练过程中,采用RFE特征选择方法,从所有输入的成矿因子中选择出对矿床(点)能正确预测的重要因子,以提供对输入模型的成矿因子的客观评价.通过对滇东南地区金矿预测的实践表明,采用SVM-RFE技术从原始10类成矿因子中自动选择6类进行预测的精度从68.42%提高到94.74%,并且得到该区域进行矿产预测的成矿因子重要性依次是:Au异常、As异常、侵入岩、下三叠统与中三叠统之间的平行不整合面、上二叠统与三叠系的平行不整合面、断裂交汇点密度、石炭系和下二叠统间的平行不整合面、中上泥盆统和石炭系间的平行不整合面、Sb异常和Hg异常,从中选取前6类成矿因子进行SVM训练得到的预测精度最高.这一结论可为在该区域进行矿产预测的资料选取,以及对成矿因子的理解提供支持.Because of the complexity and uncertainty of mineral geological information,it is a difficult task to model the distribution of mineral recourses by using parametric model.Non-linear modeling techniques such as artificial neural network（ANN）,support vector machine（SVM） etc.provide a promising mean to handle such kind of information,without acquiring exact relationship between mineral geological information and mineral deposits.In this non-linear approach,all mineral deposits and non-deposits are trained/validated/tested in a ＂black box＂-like classifier.Although a high predict accuracy can be reached if the classifier is trained properly,it is still hard to obtain classify rules,which indicate the preferable metallogenic factors from geological information because of the non-linear structural characteristic the classifier.In this paper,a technique called support vector machine based recursive feature elimination,or SVM-RFE is used to rank all input features in SVM.An experiment of SVM-RFE is conducted on gold prospectivity analysis in southeast Yunnan shows this technique could improve the predict accuracy from 68.42% to 94.74% by shrinking 10 input features to 6.The preferable rank of all 10 features calculated by SVM-RFE is Au abnormity,As abnormity,intrusive rock,parallel unconformity between Lower and Middle Triassic system,parallel unconformity between Upper Permian and Triassic system,density of faults intersection,parallel unconformity between Carboniferous and Permian system,parallel unconformity between Devonian and Carboniferous system,Sb abnormity,Hg abnormity;and the former 6 features give the best predict accuracy.This rank benefits to selection and understanding metallogenic factors in this study area.

关 键 词：特征选择 支持向量机 迭代特征消去 金矿 滇东南 

分 类 号：P628[天文地球—地质矿产勘探]