Scholarlink

<p>Accurate and timely landslide detection from remote sensing data remains challenging due to limited spatial resolution and inadequate integration of diverse geospatial information. To address these issues, this study presents a two-stage deep learning framework that couples super-resolution reconstruction with multi-source feature fusion classification. Firstly, a meta-learning-enhanced super-resolution generative adversarial network (Meta-ESRGAN) is employed to enhance the spatial feature details of optical imagery, thereby recovering fine-scale morphological characteristics that are critical for accurate landslide detection. Subsequently, a multi-source deep feature fusion classification network (ResNeSt-MD) incorporating depthwise separable convolutions (DSC) and multi-scale contextual learning, is designed to integrate the super-resolved optical data with multiple auxiliary data. The effectiveness of the proposed framework was validated using a public benchmark dataset (Bijie region, Guizhou Province, China) and a real-world case study (Nianbo area, Qinghai Province, China). Results demonstrate that the framework consistently outperforms baseline models such as ResNet50 and DensetNet121, achieving precision of 94.57% (Bijie) and 88.24% (Nianbo) respectively. Additionally, visual explanations based on Grad-CAM +  + confirm the interpretability of the model, highlighting its capability to accurately identify critical terrain features associated with landslides. This study focuses exclusively on the spatial detection of landslides using optical satellite imagery, without addressing their temporal evolution. The proposed two-stage framework enhances geomorphic edge and texture cues from optical satellite imagery, improving the reliability of inventory-based mapping.</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

Learning to detect landslides from multi-source remote sensing data with super-resolution reconstruction and deep feature fusion

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Dan Liu,
  • Keyi Zhou,
  • YuanYuan Liu

摘要

Accurate and timely landslide detection from remote sensing data remains challenging due to limited spatial resolution and inadequate integration of diverse geospatial information. To address these issues, this study presents a two-stage deep learning framework that couples super-resolution reconstruction with multi-source feature fusion classification. Firstly, a meta-learning-enhanced super-resolution generative adversarial network (Meta-ESRGAN) is employed to enhance the spatial feature details of optical imagery, thereby recovering fine-scale morphological characteristics that are critical for accurate landslide detection. Subsequently, a multi-source deep feature fusion classification network (ResNeSt-MD) incorporating depthwise separable convolutions (DSC) and multi-scale contextual learning, is designed to integrate the super-resolved optical data with multiple auxiliary data. The effectiveness of the proposed framework was validated using a public benchmark dataset (Bijie region, Guizhou Province, China) and a real-world case study (Nianbo area, Qinghai Province, China). Results demonstrate that the framework consistently outperforms baseline models such as ResNet50 and DensetNet121, achieving precision of 94.57% (Bijie) and 88.24% (Nianbo) respectively. Additionally, visual explanations based on Grad-CAM +  + confirm the interpretability of the model, highlighting its capability to accurately identify critical terrain features associated with landslides. This study focuses exclusively on the spatial detection of landslides using optical satellite imagery, without addressing their temporal evolution. The proposed two-stage framework enhances geomorphic edge and texture cues from optical satellite imagery, improving the reliability of inventory-based mapping.