Scholarlink

<p>With rapid urbanization, frequent heavy rainfall, and earthquakes, the need for a comprehensive global dataset for automated landslide mapping (LM) has become increasingly urgent. Existing datasets are limited in geographic coverage and suffer from severe class imbalance, where the number of landslide pixels is disproportionately small compared to non-landslide pixels, leading to reduced model’s performance. This study introduces the Medium Resolution Global Sentinel Landslide Dataset (MRGSLD), encompassing 21 diverse regions worldwide with varying triggering factors, lithologies and topographies. To address the dataset’s inherent imbalance and enhance the detection of landslides in remote sensing imagery, we propose a novel approach combining multiple fusion synthetic minority oversampling technique (MOF-SMOTE) with a multispectral feature attention module (MSFAM). Our approach augments the dataset through synthetic generation of underrepresented landslide pixels with MOF-SMOTE and integrates MS-FAM block into a ResAttUnet backbone to capture critical spectral features across different bands. This methodology significantly improves landslide segmentation accuracy by balancing the dataset and prioritizing relevant landslide information during downsampling. The proposed MSFAM-ResAttUnet model, when coupled with MOF-SMOTE, achieved an F1-score of 79.82, improving the benchmark model’s precision and F1-score by 28.4% and 9.3%, respectively, while MOF-SMOTE contributed an additional F1-score amelioration of 23.5–51.9%. Testing on four unseen landslide events demonstrated that the suggested method is robust, yielding results analogous to those on the MRGSLD test set, confirming the generalizability and reliability of the dataset.</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

Advancing Global Landslide Segmentation: A Coupled Multispectral Attention and Data Augmentation Approach Using the Novel MRGSLD Dataset

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • Ghislain Franck Emani,
  • Weiya Xu,
  • Kanon Guédet Guédé,
  • Firdawus Ssemugga Nattabi,
  • Olive Mekontchou Yemele

摘要

With rapid urbanization, frequent heavy rainfall, and earthquakes, the need for a comprehensive global dataset for automated landslide mapping (LM) has become increasingly urgent. Existing datasets are limited in geographic coverage and suffer from severe class imbalance, where the number of landslide pixels is disproportionately small compared to non-landslide pixels, leading to reduced model’s performance. This study introduces the Medium Resolution Global Sentinel Landslide Dataset (MRGSLD), encompassing 21 diverse regions worldwide with varying triggering factors, lithologies and topographies. To address the dataset’s inherent imbalance and enhance the detection of landslides in remote sensing imagery, we propose a novel approach combining multiple fusion synthetic minority oversampling technique (MOF-SMOTE) with a multispectral feature attention module (MSFAM). Our approach augments the dataset through synthetic generation of underrepresented landslide pixels with MOF-SMOTE and integrates MS-FAM block into a ResAttUnet backbone to capture critical spectral features across different bands. This methodology significantly improves landslide segmentation accuracy by balancing the dataset and prioritizing relevant landslide information during downsampling. The proposed MSFAM-ResAttUnet model, when coupled with MOF-SMOTE, achieved an F1-score of 79.82, improving the benchmark model’s precision and F1-score by 28.4% and 9.3%, respectively, while MOF-SMOTE contributed an additional F1-score amelioration of 23.5–51.9%. Testing on four unseen landslide events demonstrated that the suggested method is robust, yielding results analogous to those on the MRGSLD test set, confirming the generalizability and reliability of the dataset.