Scholarlink

<p>Landslides are among the most destructive natural hazards in mountainous regions, posing significant threats to lives, infrastructure, and the environment. Landslide segmentation is a critical task for disaster management, requiring robust models to handle complex terrain and environmental variations. We propose SlideNet, a novel deep learning architecture that integrates wavelet-based convolutions, adaptive fusion mechanisms, and attention-driven feature aggregation for precise segmentation of landslide regions. By leveraging wavelet transforms for multi-scale feature extraction and a tailored loss function, our approach outperforms standard segmentation models, particularly in handling the irregular boundaries and diverse textures inherent in landslide imagery. This paper details the architecture, training pipeline, and implementation specifics, with a focus on its applicability to geospatial data. SlideNet demonstrates strong performance on the LDHR and Bijie datasets, achieving state-of-the-art scores across key metrics. Qualitative and Grad-CAM visualizations confirm the model’s accuracy and interpretability. Some limitations remain in visually ambiguous regions like bare soil or roads, indicating directions for future improvement. The code for the model is available on <a href="https://github.com/asadidraco/Slidenet.git">https://github.com/asadidraco/Slidenet.git</a>.</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

Slidenet: a wavelet-enhanced architecture for landslide segmentation

  • 期刊论文

  • 发表日期：

    2025年12月14日

查看全文
  • Sonam Lhamu Bhutia,
  • Samarjeet Borah,
  • Aaditya Lochan Sharma,
  • Palash Ghosal

摘要

Landslides are among the most destructive natural hazards in mountainous regions, posing significant threats to lives, infrastructure, and the environment. Landslide segmentation is a critical task for disaster management, requiring robust models to handle complex terrain and environmental variations. We propose SlideNet, a novel deep learning architecture that integrates wavelet-based convolutions, adaptive fusion mechanisms, and attention-driven feature aggregation for precise segmentation of landslide regions. By leveraging wavelet transforms for multi-scale feature extraction and a tailored loss function, our approach outperforms standard segmentation models, particularly in handling the irregular boundaries and diverse textures inherent in landslide imagery. This paper details the architecture, training pipeline, and implementation specifics, with a focus on its applicability to geospatial data. SlideNet demonstrates strong performance on the LDHR and Bijie datasets, achieving state-of-the-art scores across key metrics. Qualitative and Grad-CAM visualizations confirm the model’s accuracy and interpretability. Some limitations remain in visually ambiguous regions like bare soil or roads, indicating directions for future improvement. The code for the model is available on https://github.com/asadidraco/Slidenet.git.