| 习泳,左鑫.基于领域自适应的胶东型金矿床三维成矿预测建模研究[J].矿产勘查,2026,17(4):635-651 |
| 基于领域自适应的胶东型金矿床三维成矿预测建模研究 |
| Research on 3D mineralization prediction modeling of Jiaodong-type gold deposits based on domain adaptation |
| 投稿时间：2025-08-26 |
| DOI：10.20008/j.kckc.202604006 |
| 中文关键词: 三维成矿预测 胶东型金矿床 迁移学习 |
| 英文关键词: 3D mineralization prediction Jiaodong-type gold deposits transfer learning |
| 基金项目:本文受地球深部探测与矿产资源勘查国家科技重大专项(2025ZD1010905)资助。 |
|
| 摘要点击次数: 186 |
| 全文下载次数: 56 |
| 中文摘要: |
| 针对深度学习方法在目标域标签稀缺时预测能力下降的问题，本文基于迁移学习思想，构建了注意力领域对抗网络模型(DANN-CTA)，对跨矿床的胶东型金矿床开展三维成矿预测建模。该模型引入梯度反转层和域判别器，在对抗训练过程中学习域不变特征，有效减小源域与目标域控矿特征分布差异，提升模型在目标域上的泛化能力。同时，三重注意力机制在空间、通道和特征层次 3个维度上动态分配权重，聚焦关键控矿因子，增强模型的特征表达能力。以招平断裂带上的夏甸和大尹格庄金矿床作为研究对象，实验结果表明，DANN-CTA模型在跨矿床预测中表现出显著优势，其 AUC值为 0.890，明显优于卷积神经网络(CNN，0.744)和随机森林(RF，0.791)，显著降低了源域与目标域控矿特征的分布差异，构建了可迁移性的胶东型金矿三维成矿预测模型，为勘查程度较低区域的三维成矿预测提供基础模型。 |
| 英文摘要: |
| To address the degradation of predictive performance in deep learning models under conditions oflimited labeled data in the target domain, this study proposes an attention-enhanced domain adversarial neural net.work (DANN-CTA) based on transfer learning for three-dimensional metallogenic prediction of Jiaodong-type golddeposits across multiple ore fields. The model integrates a gradient reversal layer and a domain discriminator to con. duct adversarial training, thereby learning domain-invariant feature representations that effectively reduce the distri.butional discrepancy of ore-controlling features between source and target domains and improve generalization capa.bility. In addition, a triple attention mechanism dynamically allocates weights across spatial, channel, and feature-level dimensions, enabling the network to focus on critical ore-controlling factors and enhance feature representa.tion.Using the Xiadian and Dayingezhuang gold deposits along the Zhaoping Fault Zone as case studies, experimen.tal results demonstrate that the proposed DANN-CTA model achieves superior cross-deposit predictive perfor.mance, attaining an AUC of 0.890, significantly outperforming the convolutional neural network (CNN, 0.744) andrandom forest (RF, 0.791) models. These results confirm that DANN-CTA effectively mitigates domain discrepan.cies in metallogenic features and establishes a transferable 3D metallogenic prediction framework for Jiaodong-typegold deposits, providing a robust methodological foundation for metallogenic prediction in underexplored regions. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |