Scholarlink

<p>Landslides are among the most destructive geological hazards, posing severe threats to human lives and property. Efficient and accurate identification is therefore essential for disaster prevention and emergency management. To address the inefficiency of conventional deep learning models in large-scale detection of newly occurred landslides from remote sensing imagery, this study proposes a spatial multi-scale collaborative framework that couples the LGBM with the YOLO deep learning model. The Yichang–Fuling region of China was chosen as the study area, with newly occurred landslides as detection targets. Ten conditioning factors (i.e., elevation, slope, aspect, profile curvature, slope morphology, topographic wetness index, distance to faults, distance to rivers, land use, and annual precipitation) were integrated, while 334 landslides were manually interpreted to construct the reference dataset. A susceptibility aggregation strategy then constrained the detection domain by aggregating 30&#xa0;m cells into 900&#xa0;m units, reducing the detection area by 89.8% while still covering 81.4% of landslides, thereby improving efficiency without compromising accuracy. Results show that the proposed framework achieved 85.6% precision, significantly surpassing the full-scale baseline (41.5%). While the global recall (80.2%) was lower than the baseline (96.4%) due to spatial constraints, the model maintained a high recall of 98.5% within the aggregated susceptibility zones. SHAP-based analysis further identified annual precipitation, slope, aspect, and land use as dominant drivers at both regional and susceptibility-aggregation scales, offering targeted insights for disaster prevention across spatial levels. This study builds a hierarchical framework from susceptibility prediction to object-based detection, emphasizing cross-scale factor consistencies and discrepancies, and offering a pathway from regional early warning to precise local landslide control.</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

Research on a spatial multi-scale detection method for newly occurred landslides by coupling LGBM and YOLO

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Xianyu Yu,
  • Zexuan Wang

摘要

Landslides are among the most destructive geological hazards, posing severe threats to human lives and property. Efficient and accurate identification is therefore essential for disaster prevention and emergency management. To address the inefficiency of conventional deep learning models in large-scale detection of newly occurred landslides from remote sensing imagery, this study proposes a spatial multi-scale collaborative framework that couples the LGBM with the YOLO deep learning model. The Yichang–Fuling region of China was chosen as the study area, with newly occurred landslides as detection targets. Ten conditioning factors (i.e., elevation, slope, aspect, profile curvature, slope morphology, topographic wetness index, distance to faults, distance to rivers, land use, and annual precipitation) were integrated, while 334 landslides were manually interpreted to construct the reference dataset. A susceptibility aggregation strategy then constrained the detection domain by aggregating 30 m cells into 900 m units, reducing the detection area by 89.8% while still covering 81.4% of landslides, thereby improving efficiency without compromising accuracy. Results show that the proposed framework achieved 85.6% precision, significantly surpassing the full-scale baseline (41.5%). While the global recall (80.2%) was lower than the baseline (96.4%) due to spatial constraints, the model maintained a high recall of 98.5% within the aggregated susceptibility zones. SHAP-based analysis further identified annual precipitation, slope, aspect, and land use as dominant drivers at both regional and susceptibility-aggregation scales, offering targeted insights for disaster prevention across spatial levels. This study builds a hierarchical framework from susceptibility prediction to object-based detection, emphasizing cross-scale factor consistencies and discrepancies, and offering a pathway from regional early warning to precise local landslide control.