基于残差图神经网络的成矿远景区预测研究

Research on Prediction of Mineralized Prospective Areas Based on Residual Graph Neural Network

摘要: 研究针对地球化学元素数据的成矿远景区预测问题，提出了一种基于残差图神经网络的深度学习框架。文章以广东省庞西垌研究区作为案例研究对象，针对地质数据稀缺、数据不平衡和深度学习模型构建难度等问题，文章采取了以下关键步骤：首先，对地球化学元素数据进行了“去闭合化”处理，以适应后续的分析；其次，针对矿区样本不足的问题，文章引入了生成对抗网络来进行数据增强，并证明了其有效性；文章提出了一种自适应阈值的皮尔森相关系数方法，将地球化学元素数据构建为图数据；最后，文章提出一种基于残差图神经网络模型，对数据进行特征提取和分类。实验结果与传统机器学习方法和其他图神经网络方法相比，文章方法在成矿远景区预测任务中表现出显著的优势。

Abstract: This study proposes a deep learning framework based on residual graph neural network for prediction of mineralized prospective areas using geochemical element data. This paper takes the Pangxidong research area in Guangdong Province as the case study object. In response to issues such as scarce geological data, imbalanced data, and difficulty in constructing deep learning models, this paper has taken the following key steps: first, this paper has “de closed” the geochemical element data to adapt to subsequent analysis; Secondly, to address the issue of insufficient samples in mining areas, this paper introduced Generative Adversarial Networks (GAN) for data augmentation and proved their effectiveness; This paper proposes an adaptive threshold Pearson correlation coefficient method to construct geochemical element data into graph data; Finally, this paper proposes a residual graph based neural network model for feature extraction and classification of data. Compared with traditional machine learning methods and other graph neural network methods, our method shows significant advantages in prediction of mineralized prospective areas.