Scholarlink

<p>This paper presents a deep ensemble learning approach for enhancing mineral prospectivity mapping (MPM) for porphyry copper deposits in the western Chagai belt, southwestern Pakistan, by integrating geophysical, geochemical, remote sensing, and geological datasets. The approach utilizes a spectral–spatial convolutional neural network (SSCNN), theoretically motivated by the coupled spectral–spatial nature of mineralization-related geospatial patterns, to extract high-dimensional features from multi-source geodata. In addition, a mineralization-based data augmentation strategy, incorporating geologically constrained spectral–spatial replacement to generate synthetic samples reflecting mineralization-favorable conditions, is employed to alleviate data imbalance and limited sample size. The ensemble of four SSCNN models achieved robust and consistent performance across four independent train–test splits, using a limited subset of known deposits for training, with F1-scores exceeding 0.94 and peaking above 0.98. The model-predicted prospectivity exhibits strong spatial correspondence with known deposits and exploration targets delineated using knowledge-driven approaches in the previous studies, particularly highlighting high-potential zones in the southeastern Reko Diq area. Field validation through rock outcrop surveys and handheld X-ray fluorescence confirmed copper and iron mineralization at eight newly identified locations. The SSCNN architecture effectively captures high-dimensional spatial and spectral features while mitigating overfitting, and the mineralization-based augmentation technique further improves model sensitivity to small-scale mineralized zones. This ensemble strategy addresses key challenges in MPM, such as limited sample size, class imbalance, data noise, and uncertainty in sample selection, while also significantly improving model generalization and robustness. Feature importance analysis revealed strong concordance between model-driven predictions and established geological knowledge, underscoring the interpretability and reliability of the method. Overall, this approach offers a powerful, data-driven framework for guiding mineral exploration in complex geological terrains and demonstrates its potential for broader application in mineralized belts globally.</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

Deep Ensemble Learning and Mineralization-Based Data Augmentation for Mineral Prospectivity Mapping of Porphyry Copper in the Western Chagai Belt, Pakistan

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Lujun Lin,
  • Lei Liu,
  • Hongrui Zhang,
  • Yasir Shaheen Khalil,
  • Jun Hong,
  • Huishan Zhang

摘要

This paper presents a deep ensemble learning approach for enhancing mineral prospectivity mapping (MPM) for porphyry copper deposits in the western Chagai belt, southwestern Pakistan, by integrating geophysical, geochemical, remote sensing, and geological datasets. The approach utilizes a spectral–spatial convolutional neural network (SSCNN), theoretically motivated by the coupled spectral–spatial nature of mineralization-related geospatial patterns, to extract high-dimensional features from multi-source geodata. In addition, a mineralization-based data augmentation strategy, incorporating geologically constrained spectral–spatial replacement to generate synthetic samples reflecting mineralization-favorable conditions, is employed to alleviate data imbalance and limited sample size. The ensemble of four SSCNN models achieved robust and consistent performance across four independent train–test splits, using a limited subset of known deposits for training, with F1-scores exceeding 0.94 and peaking above 0.98. The model-predicted prospectivity exhibits strong spatial correspondence with known deposits and exploration targets delineated using knowledge-driven approaches in the previous studies, particularly highlighting high-potential zones in the southeastern Reko Diq area. Field validation through rock outcrop surveys and handheld X-ray fluorescence confirmed copper and iron mineralization at eight newly identified locations. The SSCNN architecture effectively captures high-dimensional spatial and spectral features while mitigating overfitting, and the mineralization-based augmentation technique further improves model sensitivity to small-scale mineralized zones. This ensemble strategy addresses key challenges in MPM, such as limited sample size, class imbalance, data noise, and uncertainty in sample selection, while also significantly improving model generalization and robustness. Feature importance analysis revealed strong concordance between model-driven predictions and established geological knowledge, underscoring the interpretability and reliability of the method. Overall, this approach offers a powerful, data-driven framework for guiding mineral exploration in complex geological terrains and demonstrates its potential for broader application in mineralized belts globally.