【作者】 柏坚；

【导师】 成秋明；

【作者基本信息】 中国地质大学（北京） ， 地图制图学与地理信息工程， 2010， 博士

【摘要】 许多线性的或非线性的数学地质模型应用于矿产资源预测,来估算识别标志与预测目标矿产之间的关系。其中,选择适合的数学地质模型来达到有效的预测目的非常关键。由于地学问题本身的复杂性,传统数理统计模型在研究地学问题时具有较大局限性。当前,非线性理论与方法在资源预测与评价中的应用使矿产资源预测研究进入一个新领域。人工神经网络(Aritifical Neural Network. ANN)是非线性的复杂的动力学系统,应用到成矿预测领域,有助于深刻理解成矿系统的非线性动力学行为,更精确地再现成矿系统演化过程,对矿产资源进行更为准确的预测。另一类非线性模型——支持向量机(Support Vector Machine, SVM),从20世纪90年代以来广受重视,因其能在高维数据的有限样本情况下达到很好的适应及分类推广能力。虽然非线性模型(如ANN和SVM)在经过合理的训练后,能够得到较高的预测精度,但由于其分类过程的非线性性,分类规则不易于理解,难以从预测过程中认识到适合于特定矿产资源预测的成矿因子的知识。本文以滇东南金矿预测为研究实例,探讨非线性模型在矿产预测中的应用。另外,还研究了在预测建模前的地化数据处理和预测建模后的成矿因子的选择。本文的主要研究成果包括:(1)针对地球化学异常阈值选择这一难点,研究了多重分形分析方法,应用半径-面金属量(r-P)方法进行地球化学奇异值制图。该方法强调了数据在统计及空间域上的特征,用以反映地球化学数据集的多种类群,用来识别和圈定异常。(2)本文采用了径向基神经网络模型(Radial Basis Function Neural Network, RBFNN)对成矿地质条件复杂的滇东南地区金矿开展成矿预测,研究结果表明,该模型能快速准确的获取成矿潜力信息。(3)针对非线性分析建模(以SVM为例)难以获得分类规则(成矿条件的知识)的缺点,本文在SVM模型的训练过程中,采用RFE(Recursive Feature Elimination, RFE)特征选择方法,从各种输入的成矿因子中挖掘出对矿床(点)正确预测的重要性排序,以提供对输入模型的成矿条件的客观评价。更多还原

【Abstract】 Large numbers of mathematic geology models have been applied to mineral resources prediction and evaluation. In general, these models calculate and identify the relationships between mineral indicators and mineral targets with certain mathematical function (linear or non-linear). Therefore, it is important to select a compatible mathematic geology model to predict targets effectively. The built-in complication of geological subject limits the application of conventional mathematical statistics. The non-linear theories and methods show advantages in mineral resources prediction and evaluation. ANN (Aritifical Neural Network), a complex of non-linear system, can help to profoundly understand the non-linear process of mineralization, and then predict mineral resources more accurately. An other non-linear model, SVM (Support Vector Machine) has good adaptability and classification efficiency with limited samples of high dimension data. Although a high predict accuracy can be reached with non-linear modeling techniques, it is still hard to obtain classify rules, which indicate the preferable metallogenic factors from geological information.In this paper, the application of non-linear models to mineral resources prediction was studied with example of gold deposits prediction in Southeast Yunnan. Furthermore, the geochemical data processing method and metallogenic factors selection were studied before and in modeling. The main achievements include: (1) A geochemical anomaly identifying method, r-P model, is studied with multifractal analysis, and applied successfully to geochemical study in Southeast Yunnan. (2) RBFNN model is employed to gold prospectivity mapping in Southeast Yunnan. Experimental results show that the model can quickly obtain the probability of gold deposits. (3) For solving the shortcoming that it is difficult to get classification rules (metallogenic factor knowledge) in non-linear modeling (takes SVM as an example), this paper uses a technique called support vector machine based recursive feature elimination, or SVM-RFE to rank all input features in SVM, then values metallogenic factors in Southeast Yunnan.更多还原