Scholarlink

<p>Accurately assessing seismic hazard-prone environments in complex tectonic regions faces multiple challenges, particularly in distinguishing physical drivers from environmental proxies and mitigating spatial overfitting due to intricate interactions between geological and environmental factors. Taking Myanmar as a case study, this research proposes an integrated framework fusing multi-source geospatial data with interpretable ensemble learning to map catalog-conditioned seismic hazard-prone environments based on historical data (1970–2024). Unlike traditional perspectives that treat surface factors as direct causal drivers, we reinterpret variables such as elevation and slope as geomorphic phenotypic signatures that co-vary with seismic activity. To eliminate data leakage caused by spatial autocorrelation, a rigorous Spatial Block Cross-Validation (SBCV) strategy was implemented. Results indicate that compared to the overly optimistic outcome yielded by random splitting (AUC = 0.948), SBCV provides a more robust and conservative performance assessment (AUC = 0.890), effectively confirming the model’s generalization capability across unseen geographic units. Among the five compared models, the Gradient Boosting Decision Tree (GBDT) demonstrated superior performance. To resolve the “black-box” nature of machine learning, we integrated micro-scale SHAP values with macro-scale GeoDetector analysis. Both validation methods consistently identified Elevation, Slope, and Distance to Active Faults as core discriminators (cumulative contribution &gt; 60%), revealing that high-susceptibility zones are primarily characterized by high-relief terrain signatures that spatially coincide with active tectonic belts. Furthermore, by overlaying Nighttime Light (NTL) data, the study identified a significant risk overlap of “high-hazard environments and high exposure” within the Mandalay-Sagaing urban corridor. This study provides a scientifically rigorous and spatially validated paradigm for seismic environmental pattern recognition in data-scarce regions, offering an actionable scientific basis for regional disaster risk governance.</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

Spatially validated assessment of seismic hazard-prone environments in Myanmar using multi-source data and interpretable machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月6日

查看全文
  • Junjie Ning,
  • Zhenhui Sun,
  • Yufan Wang,
  • Ying Xu,
  • Chen Zhang

摘要

Accurately assessing seismic hazard-prone environments in complex tectonic regions faces multiple challenges, particularly in distinguishing physical drivers from environmental proxies and mitigating spatial overfitting due to intricate interactions between geological and environmental factors. Taking Myanmar as a case study, this research proposes an integrated framework fusing multi-source geospatial data with interpretable ensemble learning to map catalog-conditioned seismic hazard-prone environments based on historical data (1970–2024). Unlike traditional perspectives that treat surface factors as direct causal drivers, we reinterpret variables such as elevation and slope as geomorphic phenotypic signatures that co-vary with seismic activity. To eliminate data leakage caused by spatial autocorrelation, a rigorous Spatial Block Cross-Validation (SBCV) strategy was implemented. Results indicate that compared to the overly optimistic outcome yielded by random splitting (AUC = 0.948), SBCV provides a more robust and conservative performance assessment (AUC = 0.890), effectively confirming the model’s generalization capability across unseen geographic units. Among the five compared models, the Gradient Boosting Decision Tree (GBDT) demonstrated superior performance. To resolve the “black-box” nature of machine learning, we integrated micro-scale SHAP values with macro-scale GeoDetector analysis. Both validation methods consistently identified Elevation, Slope, and Distance to Active Faults as core discriminators (cumulative contribution > 60%), revealing that high-susceptibility zones are primarily characterized by high-relief terrain signatures that spatially coincide with active tectonic belts. Furthermore, by overlaying Nighttime Light (NTL) data, the study identified a significant risk overlap of “high-hazard environments and high exposure” within the Mandalay-Sagaing urban corridor. This study provides a scientifically rigorous and spatially validated paradigm for seismic environmental pattern recognition in data-scarce regions, offering an actionable scientific basis for regional disaster risk governance.