Scholarlink

<p>Landslides pose significant threats to biodiversity, life, and infrastructure in mountainous regions, making susceptibility mapping essential for disaster risk reduction. This study presents a comprehensive multi-model comparison of statistical, machine learning, deep learning, and ensemble approaches for landslide susceptibility mapping in the eastern Chure region of Nepal, characterized by fragile Sub-Himalayan lithology, intense monsoon precipitation, and expanding infrastructure. Ten models were evaluated: Frequency Ratio (FR), Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), XGBoost, LightGBM, Artificial Neural Network (ANN), Deep Neural Network (DNN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). Twelve conditioning factors were analyzed, including soil type, topographic wetness index, elevation, slope, stream power index, aspect, lithology, distance to roads, land use, distance to rivers, and curvature. The dataset included 123 training and 53 testing landslide locations with balanced non-landslide samples. Random Forest showed the best performance (Test AUC = 0.928; Accuracy = 0.838), followed by XGBoost (Test AUC = 0.881) and LightGBM (Test AUC = 0.844). Deep learning models showed lower performance (ANN ≈ 0.773; CNN ≈ 0.766; DNN ≈ 0.765), while LSTM produced the weakest result (Test AUC = 0.622). The Top 3 ensemble (RF, XGBoost, LightGBM) produced coherent spatial patterns and strong class separability. Classification of ensemble outputs into five susceptibility zones indicated that 22.7% of the watershed falls within High and Very High susceptibility classes. These results demonstrate that ensemble machine learning approaches provide robust landslide susceptibility assessments and can support land-use planning and disaster risk management in landslide-prone regions.</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 multi-model statistical, machine learning, and deep learning framework for landslide susceptibility in Nepal’s mid-hills

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月29日

查看全文
  • Lalit Pathak

摘要

Landslides pose significant threats to biodiversity, life, and infrastructure in mountainous regions, making susceptibility mapping essential for disaster risk reduction. This study presents a comprehensive multi-model comparison of statistical, machine learning, deep learning, and ensemble approaches for landslide susceptibility mapping in the eastern Chure region of Nepal, characterized by fragile Sub-Himalayan lithology, intense monsoon precipitation, and expanding infrastructure. Ten models were evaluated: Frequency Ratio (FR), Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), XGBoost, LightGBM, Artificial Neural Network (ANN), Deep Neural Network (DNN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). Twelve conditioning factors were analyzed, including soil type, topographic wetness index, elevation, slope, stream power index, aspect, lithology, distance to roads, land use, distance to rivers, and curvature. The dataset included 123 training and 53 testing landslide locations with balanced non-landslide samples. Random Forest showed the best performance (Test AUC = 0.928; Accuracy = 0.838), followed by XGBoost (Test AUC = 0.881) and LightGBM (Test AUC = 0.844). Deep learning models showed lower performance (ANN ≈ 0.773; CNN ≈ 0.766; DNN ≈ 0.765), while LSTM produced the weakest result (Test AUC = 0.622). The Top 3 ensemble (RF, XGBoost, LightGBM) produced coherent spatial patterns and strong class separability. Classification of ensemble outputs into five susceptibility zones indicated that 22.7% of the watershed falls within High and Very High susceptibility classes. These results demonstrate that ensemble machine learning approaches provide robust landslide susceptibility assessments and can support land-use planning and disaster risk management in landslide-prone regions.