Scholarlink

<p>Clustered landslides triggered by extreme rainfall and major earthquakes are increasingly frequent, necessitating rapid, accurate landslide detection (LSD) tools for post-disaster response. However, traditional LSD methods encounter challenges, including limited samples in few-shot scenarios and the underexplored use of Transformer-based models. This study presents LSDFormer, a compact CNN–Transformer hybrid for cross-region LSD, and develops a globally distributed landslide dataset (GDLD) for pre-training and transfer learning (TL). LSDFormer adopts a U-shaped encoder–decoder with lightweight blocks for efficient local and global feature capture (20.0&#xa0;G FLOPs, 37.7&#xa0;M parameters). It integrates two novel blocks: Efficient-CT, which combines a lightweight MobileNetV3 block (MB3) and efficient multi-head self-attention (EMSA) to jointly capture local details and global dependencies, and Enhanced-CT, which leverages the spatial specificity of involution with multi-level involution-based spatial attention (MISA) to further enhance feature representation. To evaluate the cross-regional generalization, we constructed two target domain datasets: RLZX (rainfall-induced landslides in Zixing, China) and ELNT (earthquake-induced landslides in the Noto Peninsula, Japan). Experiments conducted on the RLZX and ELNT datasets demonstrate that LSDFormer surpasses existing state-of-the-art methods. Additionally, our TL framework enables the model to achieve high-quality results with only 10% or even less of target domain data. Our model and framework excel in rapidly detecting rainfall-induced and earthquake-induced landslides across large areas, reducing labeling needs and completing LSD tasks for approximately 1,000 km<sup>2</sup> in minutes to meet post-disaster emergency response demands, demonstrating substantial practical value. It also establishes a promising paradigm for future LSD research.</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 CNN–Transformer hybrid network for efficient cross-region landslide detection by transfer learning

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Zijin Fu,
  • Fawu Wang,
  • Junfei Zhong,
  • Filippo Catani,
  • Jie Dou,
  • Qi You,
  • Bo Zhang

摘要

Clustered landslides triggered by extreme rainfall and major earthquakes are increasingly frequent, necessitating rapid, accurate landslide detection (LSD) tools for post-disaster response. However, traditional LSD methods encounter challenges, including limited samples in few-shot scenarios and the underexplored use of Transformer-based models. This study presents LSDFormer, a compact CNN–Transformer hybrid for cross-region LSD, and develops a globally distributed landslide dataset (GDLD) for pre-training and transfer learning (TL). LSDFormer adopts a U-shaped encoder–decoder with lightweight blocks for efficient local and global feature capture (20.0 G FLOPs, 37.7 M parameters). It integrates two novel blocks: Efficient-CT, which combines a lightweight MobileNetV3 block (MB3) and efficient multi-head self-attention (EMSA) to jointly capture local details and global dependencies, and Enhanced-CT, which leverages the spatial specificity of involution with multi-level involution-based spatial attention (MISA) to further enhance feature representation. To evaluate the cross-regional generalization, we constructed two target domain datasets: RLZX (rainfall-induced landslides in Zixing, China) and ELNT (earthquake-induced landslides in the Noto Peninsula, Japan). Experiments conducted on the RLZX and ELNT datasets demonstrate that LSDFormer surpasses existing state-of-the-art methods. Additionally, our TL framework enables the model to achieve high-quality results with only 10% or even less of target domain data. Our model and framework excel in rapidly detecting rainfall-induced and earthquake-induced landslides across large areas, reducing labeling needs and completing LSD tasks for approximately 1,000 km2 in minutes to meet post-disaster emergency response demands, demonstrating substantial practical value. It also establishes a promising paradigm for future LSD research.