Scholarlink

<p>Reliable debris-flow susceptibility mapping (DFSM) hinges on both optimal topographic characterization and unbiased data sampling. This study systematically&#xa0;quantifies the effect of four DEM resolutions (6.5–90&#xa0;m) and sampling strategy uncertainty on advanced deep learning (DL) models, including convolutional neural network (CNN1D, CNN2D), recurrent neural network (RNN), and long short-term memory (LSTM). A field-based comprehensive debris flow inventory (108 debris flow watersheds) and 13 contributing factors derived from multi-resolution DEMs and remote sensing data were utilized to construct the datasets. To assess the impact of sampling on model performance, an extensive uncertainty analysis was conducted through 100 symmetrical sampling iterations (debris flow and non-debris flow), resulting in a 6.7–10.5% improvement in mean accuracy by varying sampling points. Feature importance was further examined using a Random Forest classifier and frequency ratio (FR), identifying rainfall as the most significant factor. Model performance was evaluated through multiple metrics such as accuracy, recall, Jaccard, precision, kappa, AUC, and F1-score. Among the models, LSTM consistently outperformed other compared models across all resolutions, achieving the maximum debris flow susceptibility accuracy of 0.929 and AUC of 0.973 at 12.5&#xa0;m resolution. The proposed multi-resolution DL framework provides a reproducible pathway for uncertainty-aware debris-flow susceptibility assessment in complex mountainous terrains.</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

The role of multi-resolution DEMs and sampling strategy uncertainty in deep learning-based debris flow susceptibility mapping

  • 期刊论文

  • 发表日期：

    2025年11月18日

查看全文
  • Hamza Daud,
  • Jie Dou,
  • Javed Iqbal Tanoli,
  • Nafees Ali,
  • Nangyal Ghani Khan,
  • Zilin Xiang,
  • Aonan Dong,
  • Ke Xing,
  • Hafiz Ullah,
  • Lele Zhang

摘要

Reliable debris-flow susceptibility mapping (DFSM) hinges on both optimal topographic characterization and unbiased data sampling. This study systematically quantifies the effect of four DEM resolutions (6.5–90 m) and sampling strategy uncertainty on advanced deep learning (DL) models, including convolutional neural network (CNN1D, CNN2D), recurrent neural network (RNN), and long short-term memory (LSTM). A field-based comprehensive debris flow inventory (108 debris flow watersheds) and 13 contributing factors derived from multi-resolution DEMs and remote sensing data were utilized to construct the datasets. To assess the impact of sampling on model performance, an extensive uncertainty analysis was conducted through 100 symmetrical sampling iterations (debris flow and non-debris flow), resulting in a 6.7–10.5% improvement in mean accuracy by varying sampling points. Feature importance was further examined using a Random Forest classifier and frequency ratio (FR), identifying rainfall as the most significant factor. Model performance was evaluated through multiple metrics such as accuracy, recall, Jaccard, precision, kappa, AUC, and F1-score. Among the models, LSTM consistently outperformed other compared models across all resolutions, achieving the maximum debris flow susceptibility accuracy of 0.929 and AUC of 0.973 at 12.5 m resolution. The proposed multi-resolution DL framework provides a reproducible pathway for uncertainty-aware debris-flow susceptibility assessment in complex mountainous terrains.