Scholarlink

<p>Machine learning methods for mineral prospectivity prediction, including convolutional neural networks (CNN), can effectively capture the complex nonlinear relationships between geological exploration big data and mineral prospectivity and have been widely applied. However, the generalization ability of the predictive models formed through training remains limited. Additionally, mineralization is inherently a rare event, and the scarcity of mineral deposit samples in exploration data negatively impacts prediction accuracy. To address these issues, this study integrated self-training semi-supervised techniques with interpretable deep learning methods to establish a dual-channel iterative framework. This framework enhances the scientific reliability of mineral prospectivity results by focusing on result interpretability. Class rebalancing techniques were applied to increase the number of mineral deposit samples, and the SHapley Additive exPlanations (SHAP) method was employed to calculate feature output weights. These processes were iteratively embedded into the CNN, resulting in more accurate and reliable predictions. The proposed method mitigates the negative effects of data imbalance and enhances interpretability. An experiment in the Keeryin ore concentration area of Sichuan, China, validated this approach, showing that 92% of the samples fell within high mineralization probability zones, which accounted for 9.73% of the total area. SHAP analysis elucidated the evolving contributions of key mineralization-controlling features, such as Li/La ratio and hydroxyl anomaly, which align with the regional metallogenic mechanisms.</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

A Dual-Channel Iterative Method Integrating Semi-supervised Self-Training and Interpretable Deep Learning Models for Mineral Prospectivity Prediction

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Shitao Yin,
  • Nan Li,
  • Keyan Xiao,
  • Xianglong Song,
  • Xingjie Wang

摘要

Machine learning methods for mineral prospectivity prediction, including convolutional neural networks (CNN), can effectively capture the complex nonlinear relationships between geological exploration big data and mineral prospectivity and have been widely applied. However, the generalization ability of the predictive models formed through training remains limited. Additionally, mineralization is inherently a rare event, and the scarcity of mineral deposit samples in exploration data negatively impacts prediction accuracy. To address these issues, this study integrated self-training semi-supervised techniques with interpretable deep learning methods to establish a dual-channel iterative framework. This framework enhances the scientific reliability of mineral prospectivity results by focusing on result interpretability. Class rebalancing techniques were applied to increase the number of mineral deposit samples, and the SHapley Additive exPlanations (SHAP) method was employed to calculate feature output weights. These processes were iteratively embedded into the CNN, resulting in more accurate and reliable predictions. The proposed method mitigates the negative effects of data imbalance and enhances interpretability. An experiment in the Keeryin ore concentration area of Sichuan, China, validated this approach, showing that 92% of the samples fell within high mineralization probability zones, which accounted for 9.73% of the total area. SHAP analysis elucidated the evolving contributions of key mineralization-controlling features, such as Li/La ratio and hydroxyl anomaly, which align with the regional metallogenic mechanisms.