Scholarlink

<p>Mineral prospectivity mapping is central to modern exploration, yet conventional machine learning approaches yield deterministic class probabilities that overstate confidence and ignore spatial heterogeneity. Here, a spatially aware evidential deep learning (SAEDL) framework is introduced, which integrates spatially weighted regression with evidential deep neural networks to simultaneously model mineral prospectivity and decompose uncertainty into aleatoric and epistemic components. In the SAEDL architecture, spatial adaptivity is achieved by generating location-specific regression coefficients and biases through a feed-forward network modulated by learned Gaussian kernel functions. These parameters are transformed into Dirichlet concentration parameters via softplus activation, allowing the model to estimate class probabilities and associated uncertainties. To improve interpretability, the integrated gradients method is employed to attribute the log-odds of prediction back to individual features, offering spatially resolved explanations of model behavior. The SAEDL approach is validated through a tungsten polymetallic prospecting case study in the Nanling metallogenic belt (China), a region with complex mineralization controls. The model attains area under the receiver operating characteristic (ROC) curve (AUC) values of 0.92 (training data) and 0.91 (test data), with stable, monotonic convergence of loss curves and well-calibrated uncertainties as measured by the expected calibration error and the Brier score. The SAEDL decomposes predictive uncertainty into aleatoric and epistemic components: high aleatoric uncertainty highlights intrinsic data noise, whereas elevated epistemic uncertainty pinpoints regions where additional sampling could most reduce risk. The prospectivity map and uncertainty map jointly guide risk-aware exploration, prioritizing areas of high predicted probability with low total uncertainty and identifying zones for targeted data acquisition. By unifying spatial adaptivity, transparent feature attribution, and uncertainty quantification in an efficient non-Bayesian architecture, the SAEDL provides a rigorous, risk-aware tool for regional exploration work and establishes a universal template for spatial classification tasks beyond mineral prospectivity.</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 Spatially Aware Evidential Deep Learning Framework for Mineral Prospectivity Mapping and Uncertainty Evaluation

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Yue Liu

摘要

Mineral prospectivity mapping is central to modern exploration, yet conventional machine learning approaches yield deterministic class probabilities that overstate confidence and ignore spatial heterogeneity. Here, a spatially aware evidential deep learning (SAEDL) framework is introduced, which integrates spatially weighted regression with evidential deep neural networks to simultaneously model mineral prospectivity and decompose uncertainty into aleatoric and epistemic components. In the SAEDL architecture, spatial adaptivity is achieved by generating location-specific regression coefficients and biases through a feed-forward network modulated by learned Gaussian kernel functions. These parameters are transformed into Dirichlet concentration parameters via softplus activation, allowing the model to estimate class probabilities and associated uncertainties. To improve interpretability, the integrated gradients method is employed to attribute the log-odds of prediction back to individual features, offering spatially resolved explanations of model behavior. The SAEDL approach is validated through a tungsten polymetallic prospecting case study in the Nanling metallogenic belt (China), a region with complex mineralization controls. The model attains area under the receiver operating characteristic (ROC) curve (AUC) values of 0.92 (training data) and 0.91 (test data), with stable, monotonic convergence of loss curves and well-calibrated uncertainties as measured by the expected calibration error and the Brier score. The SAEDL decomposes predictive uncertainty into aleatoric and epistemic components: high aleatoric uncertainty highlights intrinsic data noise, whereas elevated epistemic uncertainty pinpoints regions where additional sampling could most reduce risk. The prospectivity map and uncertainty map jointly guide risk-aware exploration, prioritizing areas of high predicted probability with low total uncertainty and identifying zones for targeted data acquisition. By unifying spatial adaptivity, transparent feature attribution, and uncertainty quantification in an efficient non-Bayesian architecture, the SAEDL provides a rigorous, risk-aware tool for regional exploration work and establishes a universal template for spatial classification tasks beyond mineral prospectivity.