基于ROC曲线分析理论的矿产预测与效果评价的通用效益-代价模型研究及应用示范

The modern theory and method system of mineral prediction lacks a universal model and criterion for mineral prediction performance assessment. The ROC curve analysis theory and its applications in machine learning and medical diagnosis provide a useful reference for developing the universal model and criterion for the mineral prediction performance assessment. For this reason, our research group plan to introduce the ROC curve analysis theory into the research field of the mineral prediction to develop an ROC curve analysis theory based universal benefit-cost model for the mineral prediction and prediction performance assessment. In the new model, the overall goal of mineral prediction is defined quantitatively as the universal mineral prediction criterion called minimizing costs while maximizing benefits. This criterion can be used to properly deal with a set of the key scientific problems such as metallogenic evidence selection, the optimal segmentation of continuous metallogenic evidences, mineral target prediction, prediction performance assessment, and so on, and further make the mineral prediction procedure become an automated process. The multisource geo-data including geological, geochemical, geophysical, and remote sensing data serve as the basic data sources for the mineral prediction. Application examples of the universal benefit-cost model for the mineral prediction and prediction performance assessment will be established in the two application demonstration areas. This research can expand the application fields of the ROC curve analysis theory and promote the development of the modern theory and method system of the mineral prediction. The expected achievements have both important theoretical significance and practical popularization value.

当代矿产预测理论与方法体系缺少矿产预测效果评价的通用模型和准则。ROC曲线分析理论及其在机器学习和医疗诊断领域的应用研究成果为研制矿产预测效果评价的通用模型和准则提供了一种有益的借鉴。鉴于此，项目组拟将ROC曲线分析理论引入矿产预测研究领域，研制基于ROC曲线分析理论的矿产预测和预测效果评价的通用效益－代价模型。新模型把矿产预测的总体目标量化为"最小化代价－最大化效益"的矿产预测通用准则，用来解决找矿证据优选、连续型找矿证据最优离散化、矿产靶区预测和预测效果评价等一系列关键科学问题，使矿产预测过程实现自动化。以地物化遥多源地学数据为基本数据源，拟在两个示范区创建通用效益－代价模型矿产预测和预测效果评价应用研究的典型范例。此项研究能够拓宽ROC曲线分析理论的应用研究领域，促进当代矿产预测理论与方法体系的发展，预期研究成果具有重要的理论意义和推广应用价值。

ROC曲线分析理论起源于二战时期的雷达目标识别性能评价。后来，该理论被广泛应用于机器学习领域的二元分类模型性能评价和医学领域的统计诊断。针对当代矿产资源评价理论与方法体系缺少矿产预测效果评价通用模型和准则的现状，项目组把ROC曲线分析理论引入矿产资源评价研究领域，借鉴ROC曲线分析理论在机器学习和医疗诊断领域的应用研究成果，研制出基于ROC曲线分析理论的矿产靶区预测及效果评价的通用效益－代价模型。该模型把矿产靶区预测的总体目标量化为“最小化代价－最大化效益”的矿产资源评价通用准则，用来解决找矿证据优选、连续型找矿证据最优离散化、矿产靶区预测及效果评价等一系列关键科学问题，从而提高矿产靶区预测过程的自动化水平。选择青海拉陵灶火地区为实验区，将研究区划分成200×187个网格统计单元（含矿单元17个），构建了基于ROC曲线分析理论的多金属矿产靶区预测及效果评价的通用效益-代价模型，用约登指数、似然比指数和提升指数表示统计单元成矿有利度，预测了多金属矿产靶区，用ROC曲线和提升指数曲线评价了矿产靶区预测效果，用AUC值评价了矿产靶区预测的总体效果，用约登指数确定了矿产靶区与非靶区的最佳分界线并圈定了最优预测靶区。效益-代价模型和证据权模型的矿产靶区预测效果的对比分析结果表明：①ROC曲线特征和AUC值揭示效益-代价模型能够很好地区分研究区含矿和非含矿统计单元；②效益-代价模型预测的最优靶区占研究区总面积为9.36~11.19%，但包含了研究区82%的已知多金属矿床（点）；③效益-代价模型优于目前广泛应用的证据权模型。此项研究拓宽了ROC曲线分析理论的应用研究领域，促进当代矿产评价理论与方法体系的发展，具有重要的理论意义和推广应用价值。