基于地质图编码及深度残差网络的找矿预测方法——以陕西石泉地区金矿为例

Ore prospecting and prediction method based on geological map encoding and deep residual networks—A case study of gold deposits in Shiquan area,Shaanxi

利用人工智能技术进行找矿预测是矿产勘查的前沿领域,解决了地质图因精度不足而无法被用于找矿预测及模型训练等问题。提出并采用ONE-HOT编码方法进行地质图网格化,结合地球化学数据,以多通道二维网格窗口数据作为输入数据,利用平移、旋转、缩放等数据增强方法进行扩充训练,采用改进的ResNet神经网络预测模型,通过大量试验与对比,优化了超参数。在石泉地区进行了找矿预测,结果表明,该方法的找矿远景区面积占比6.5%,覆盖了所有已知金矿,并预测出8处不包含已知金矿的新找矿远景区,所提出的ONE-HOT编码方法可以有效利用已有地质图数据,且ResNet神经网络预测模型预测结果明显优于典型CNN网络模型和证据权重法的预测结果。

The use of artificial intelligence technology for ore prospecting and prediction has become a frontier area in mineral exploration,solving the problem of insufficient accuracy in geological maps for prospecting and prediction and model training.In this study,the ONE-HOT encoding method is proposed and utilized to grid the geological maps.In association with geochemical data,multi-channel two-dimensional grid window data are used as input,and data augmentation methods such as translation,rotation,and scaling are employed to expand the training.An improved ResNet neural network prediction model is used and optimizes hyperparameters through numerous experiments and comparisons.Ore prospecting and prediction was conducted in the Shiquan area,and the results show that the prospective area identified by this method covers 6.5%of the entire region,including all known gold deposits.Additionally,8 new prospective areas without known gold deposits were predicted.The proposed ONE-HOT encoding method effectively utilizes existing geological map data,and the ResNet neural network prediction model performs significantly better than typical CNN network models and weight of evidence methods.