Scholarlink

<p>This research presents a new approach to predict landslide occurrences and develop susceptibility maps to support early warning and risk reduction. The proposed approach is based on enhancing the forecast accuracy of decision tree-based machine learning models, including Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LGBM), and Extreme Gradient Boosting (XGBoost), using optimization algorithms. These optimization functions include the marine predator’s algorithm (MPA), grey wolf optimization (GWO), shrike optimization algorithm (SHOA), and rhinopithecus swarm optimization algorithm (RSOA). On the other hand, stratified sampling techniques are used to develop a model capable of capturing features of an imbalanced dataset, where the number of non-landslide areas overwhelmingly exceeds that of landslide points. The best predictive model is selected through a comprehensive evaluation framework integrating ROC-AUC, AUPRC, F1-score, Cohen’s Kappa, and SCAI to address the extreme class imbalance. Statistical significance was further validated via the Chi-square test to justify the impact of swarm optimization. As a result, the CatBoost-RSOA model, combined with stratified sampling splitting techniques, is the best-performing model, achieving a test AUC of 77.7%. This is an improvement of around 7% over the physical model (70.7%) and 4% compared with the Monte Carlo simulation model (73.8%) in previous studies.</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

Enhancing the prediction of landslide susceptibility mapping by using hybrid machine learning models with imbalanced data

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Viet Dat Le,
  • Yun-Tae Kim,
  • Huu Nghia Bui,
  • My Quoc Dang,
  • Ngoc Thi Huynh,
  • Van Qui Lai

摘要

This research presents a new approach to predict landslide occurrences and develop susceptibility maps to support early warning and risk reduction. The proposed approach is based on enhancing the forecast accuracy of decision tree-based machine learning models, including Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LGBM), and Extreme Gradient Boosting (XGBoost), using optimization algorithms. These optimization functions include the marine predator’s algorithm (MPA), grey wolf optimization (GWO), shrike optimization algorithm (SHOA), and rhinopithecus swarm optimization algorithm (RSOA). On the other hand, stratified sampling techniques are used to develop a model capable of capturing features of an imbalanced dataset, where the number of non-landslide areas overwhelmingly exceeds that of landslide points. The best predictive model is selected through a comprehensive evaluation framework integrating ROC-AUC, AUPRC, F1-score, Cohen’s Kappa, and SCAI to address the extreme class imbalance. Statistical significance was further validated via the Chi-square test to justify the impact of swarm optimization. As a result, the CatBoost-RSOA model, combined with stratified sampling splitting techniques, is the best-performing model, achieving a test AUC of 77.7%. This is an improvement of around 7% over the physical model (70.7%) and 4% compared with the Monte Carlo simulation model (73.8%) in previous studies.