Scholarlink

<p>Snow avalanches are among the most destructive natural hazard in the Himalayan regions, causing significant threats to human life, infrastructure and the mountain environment. This study develops a Bayesian-optimized hybrid deep learning framework that combines CNN–LSTM and CNN–LSTM-Attention models to predict snow avalanche susceptibility with improved accuracy and interpretability. Sixteen snow avalanche influencing factors (SAIFs) covering topographic, climatic, environmental, hydrological, anthropogenic, and geological domains were selected. Bayesian optimization was employed to tune critical model hyperparameters, enhancing convergence stability and predictive performance. The results show that the CNN-LSTM-Attention model outperformed the CNN-LSTM model, achieving an AUC of 0.94 and an F1-score of 0.98 and captured a larger area (13.01%) under very high susceptibility zones compared to the CNN-LSTM model (3.68%). Further, model interpretability was enhanced through SHapley Additive exPlanations (SHAP) analysis, which identified LU/LC, NDVI, TPI, elevation, and slope as the most influential factors driving snow avalanche susceptibility. The SHAP dependence plots further revealed complex nonlinear relationships between the key predictors and model output. Overall, this study demonstrates that the integration of Bayesian optimization, attention mechanisms, and SHAP-based interpretability offers a robust, transparent, and highly effective framework for snow avalanche susceptibility modeling, providing valuable insights for hazard assessment and risk-informed land-use planning in mountainous regions.</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

An Explainable Bayesian-Optimized Hybrid Deep Learning Framework with Attention Mechanisms for Modeling Snow Avalanche Susceptibility

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Intikhab Ahmad,
  • Javed Mallick,
  • Hoang Thi Hang,
  • Mohd Rihan,
  • Asmita Bakshi

摘要

Snow avalanches are among the most destructive natural hazard in the Himalayan regions, causing significant threats to human life, infrastructure and the mountain environment. This study develops a Bayesian-optimized hybrid deep learning framework that combines CNN–LSTM and CNN–LSTM-Attention models to predict snow avalanche susceptibility with improved accuracy and interpretability. Sixteen snow avalanche influencing factors (SAIFs) covering topographic, climatic, environmental, hydrological, anthropogenic, and geological domains were selected. Bayesian optimization was employed to tune critical model hyperparameters, enhancing convergence stability and predictive performance. The results show that the CNN-LSTM-Attention model outperformed the CNN-LSTM model, achieving an AUC of 0.94 and an F1-score of 0.98 and captured a larger area (13.01%) under very high susceptibility zones compared to the CNN-LSTM model (3.68%). Further, model interpretability was enhanced through SHapley Additive exPlanations (SHAP) analysis, which identified LU/LC, NDVI, TPI, elevation, and slope as the most influential factors driving snow avalanche susceptibility. The SHAP dependence plots further revealed complex nonlinear relationships between the key predictors and model output. Overall, this study demonstrates that the integration of Bayesian optimization, attention mechanisms, and SHAP-based interpretability offers a robust, transparent, and highly effective framework for snow avalanche susceptibility modeling, providing valuable insights for hazard assessment and risk-informed land-use planning in mountainous regions.