Scholarlink

<p>The evaluation of glacier lake susceptibility and extraction is critical to under- standing and mitigating the risks posed by glacial lake outburst floods (GLOF). This paper presents a hybrid deep learning-based global-local salient mapping network (GLSM-Net) for glacier lake segmentation and GLOF risk assessment using Landsat 8 satellite imagery. The proposed model integrates global and local salient feature modules (GSFM and LSFM) to capture multi-scale contextual and fine-grained spatial features. A dual-pathway U-Net + + and SE-Net encoder-decoder architecture with feature cross-fusion blocks (FCFB) is employed for improved lake boundary delineation and risk scoring. GLOF risk scores are also generated through dense layers following feature extraction, enabling precise flood susceptibility predictions. Experimental results across the Imja Glacier, Chandra Basin, and Bhaga Glacier regions show a performance improvement over existing segmentation models, achieving 95.06% accuracy and F1 scores exceeding 93%. This framework enhances early flood risk prediction and aids disaster management efforts.</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

GLSM-Net: A Global-Local Salient Mapping Network for Glacial Lake Extraction and Risk Assessment

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Jagadeesh Thati,
  • Samit Ari

摘要

The evaluation of glacier lake susceptibility and extraction is critical to under- standing and mitigating the risks posed by glacial lake outburst floods (GLOF). This paper presents a hybrid deep learning-based global-local salient mapping network (GLSM-Net) for glacier lake segmentation and GLOF risk assessment using Landsat 8 satellite imagery. The proposed model integrates global and local salient feature modules (GSFM and LSFM) to capture multi-scale contextual and fine-grained spatial features. A dual-pathway U-Net + + and SE-Net encoder-decoder architecture with feature cross-fusion blocks (FCFB) is employed for improved lake boundary delineation and risk scoring. GLOF risk scores are also generated through dense layers following feature extraction, enabling precise flood susceptibility predictions. Experimental results across the Imja Glacier, Chandra Basin, and Bhaga Glacier regions show a performance improvement over existing segmentation models, achieving 95.06% accuracy and F1 scores exceeding 93%. This framework enhances early flood risk prediction and aids disaster management efforts.