Scholarlink

<p>The detection of geochemical anomalies plays an essential role in mineral exploration. Geochemical survey data are characterized by complex multi-scale spatial patterns, with element concentrations exhibiting significant spatial autocorrelation, heterogeneity, and local singularity. Conventional deep learning algorithms typically operate at just one fixed scale, which can make it difficult for them to handle complex data and result in the fragmented delineation of large anomalies or excessive smoothing of smaller but geologically significant anomalies. To address these limitations, GeoYOLO-GCN is developed as a novel framework that integrates the multi-scale feature pyramid, graph convolutional networks (GCN), and Transformer-based attention mechanisms of the You Only Look Once (YOLO) model. This framework extracts and integrates multi-scale geochemical features using a GCN module to effectively extract the spatial patterns of geochemical survey data and a Transformer module for adaptive cross-scale feature fusion that dynamically balances the importance of different scale patterns. Simultaneously, the focal loss mechanism guides the model to focus on rare and hard-to-classify anomaly samples, effectively mitigating the issue of data imbalance between sparse mineralization points and extensive background areas. In addition, a geologically constrained focal loss function incorporates prior geological knowledge as a soft constraint to penalize any predictions that are inconsistent with established metallogenic principles for enhancing the interpretability of the results. The effectiveness of the proposed model was validated on a stream sediment geochemical dataset from a gold deposit district in the Gansu Province of China. Results from ablation studies demonstrate the superiority of the GeoYOLO-GCN architecture with an area under the receiver operating characteristic curve of 0.960. Furthermore, the success rate curves show that this model successfully identifies 93.5% of known gold deposits within only 10.0% of the study area. Compared with the unconstrained YOLO-GCN model, the GeoYOLO-GCN model can improve identification performance and geological interpretability.</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

Geologically Constrained Multi-scale Graph Convolutional Network Modeling of Stream Sediment Geochemical Data

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Yuesen Zhang,
  • Zhiyi Chen,
  • Denghong Long,
  • Renguang Zuo

摘要

The detection of geochemical anomalies plays an essential role in mineral exploration. Geochemical survey data are characterized by complex multi-scale spatial patterns, with element concentrations exhibiting significant spatial autocorrelation, heterogeneity, and local singularity. Conventional deep learning algorithms typically operate at just one fixed scale, which can make it difficult for them to handle complex data and result in the fragmented delineation of large anomalies or excessive smoothing of smaller but geologically significant anomalies. To address these limitations, GeoYOLO-GCN is developed as a novel framework that integrates the multi-scale feature pyramid, graph convolutional networks (GCN), and Transformer-based attention mechanisms of the You Only Look Once (YOLO) model. This framework extracts and integrates multi-scale geochemical features using a GCN module to effectively extract the spatial patterns of geochemical survey data and a Transformer module for adaptive cross-scale feature fusion that dynamically balances the importance of different scale patterns. Simultaneously, the focal loss mechanism guides the model to focus on rare and hard-to-classify anomaly samples, effectively mitigating the issue of data imbalance between sparse mineralization points and extensive background areas. In addition, a geologically constrained focal loss function incorporates prior geological knowledge as a soft constraint to penalize any predictions that are inconsistent with established metallogenic principles for enhancing the interpretability of the results. The effectiveness of the proposed model was validated on a stream sediment geochemical dataset from a gold deposit district in the Gansu Province of China. Results from ablation studies demonstrate the superiority of the GeoYOLO-GCN architecture with an area under the receiver operating characteristic curve of 0.960. Furthermore, the success rate curves show that this model successfully identifies 93.5% of known gold deposits within only 10.0% of the study area. Compared with the unconstrained YOLO-GCN model, the GeoYOLO-GCN model can improve identification performance and geological interpretability.