Scholarlink

<p>Mineral prospectivity mapping (MPM), which aims to identify areas with high mineral potential, has recently attracted greater attention owing to the application of deep learning algorithms. Mineralization is a coupled event arising from multiple geological processes within specific favorable spaces and generally exhibits multiple stages and clustering distributions. Therefore, fully incorporating both the mineralization processes and their spatial distribution offers significant advantages for improving MPM. However, most existing deep learning-based MPM models tend to emphasize either mineralization processes or mineralization spatial distribution, rather than integrating both. This limitation is particularly evident in 3D MPM. This paper presents a process–space dual-branch transformer with gating fusion (PSDBTG) model that considers processes and spatial distribution of mineralization, enabling the comprehensive utilization of process- and space-related features for 3D MPM. The process-based branch utilizes evidence layers that represent the mineralization processes or exploration criteria to capture the process-related features. The space-based branch exploits the interactions among neighboring voxels to obtain local spatial features related to orebody distribution. Subsequently, a gating module is employed to adaptively integrate process- and space-related features with different emphases from the two branches. Here, the PSDBTG model was applied to conduct 3D MPM in the Lannigou gold district, China. The comparative experiments demonstrated that the PSDBTG model, which incorporates both processes and spatial distribution of mineralization, outperformed the models that consider only a single type of information in terms of prediction performance. Moreover, the gating module exhibited superiority over other feature fusion methods by adaptively regulating negative constraints derived from mineralization processes. This paper proposes a novel and transparent framework for embedding geological prior knowledge into data-driven MPM, which allows model predictions to explicitly incorporate mineralization process constraints, contributing to improved interpretability and reliability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-Branch Transformer Model Considering Processes and Spatial Distribution of Mineralization for 3D Mineral Prospectivity Mapping

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Zhongzheng Wang,
  • Gongwen Wang,
  • Zhiqiang Zhang,
  • Qingming Peng

摘要

Mineral prospectivity mapping (MPM), which aims to identify areas with high mineral potential, has recently attracted greater attention owing to the application of deep learning algorithms. Mineralization is a coupled event arising from multiple geological processes within specific favorable spaces and generally exhibits multiple stages and clustering distributions. Therefore, fully incorporating both the mineralization processes and their spatial distribution offers significant advantages for improving MPM. However, most existing deep learning-based MPM models tend to emphasize either mineralization processes or mineralization spatial distribution, rather than integrating both. This limitation is particularly evident in 3D MPM. This paper presents a process–space dual-branch transformer with gating fusion (PSDBTG) model that considers processes and spatial distribution of mineralization, enabling the comprehensive utilization of process- and space-related features for 3D MPM. The process-based branch utilizes evidence layers that represent the mineralization processes or exploration criteria to capture the process-related features. The space-based branch exploits the interactions among neighboring voxels to obtain local spatial features related to orebody distribution. Subsequently, a gating module is employed to adaptively integrate process- and space-related features with different emphases from the two branches. Here, the PSDBTG model was applied to conduct 3D MPM in the Lannigou gold district, China. The comparative experiments demonstrated that the PSDBTG model, which incorporates both processes and spatial distribution of mineralization, outperformed the models that consider only a single type of information in terms of prediction performance. Moreover, the gating module exhibited superiority over other feature fusion methods by adaptively regulating negative constraints derived from mineralization processes. This paper proposes a novel and transparent framework for embedding geological prior knowledge into data-driven MPM, which allows model predictions to explicitly incorporate mineralization process constraints, contributing to improved interpretability and reliability.