What question did this study set out to answer?

To assess the predictive performance of various machine learning models on landslide susceptibility with imbalanced data.

April 18, 2026Open Access

Tree-based and deep learning models for landslide susceptibility assessment with imbalanced data

Key Points

To assess the predictive performance of various machine learning models on landslide susceptibility with imbalanced data.
Utilized a preprocessing pipeline including Synthetic Minority Oversampling Technique and Robust Scaling.
Conducted comparative analysis of five models: RF, XGBoost, CatBoost, LightGBM, and TabNet.
Applied metrics such as AUC, accuracy, recall, F1-score, and precision for evaluation.
LightGBM achieved the highest AUC of 0.980, indicating superior predictive ability.
XGBoost had the best overall performance across combined metrics like accuracy and F1-score.
RF demonstrated strong results consistently across multiple criteria, while TabNet showed potential with a competitive recall.

Abstract

Landslides are phenomena that pose substantial risks to human life, infrastructure, and environmental sustainability, necessitating precise and reliable susceptibility assessment to guide effective mitigation efforts. However, landslide monitoring datasets often exhibit severe class imbalance due to the rarity of events, along with inconsistent feature distributions. These challenges adversely impact the predictive performance and generalizability of machine learning models. To address these issues, this study employs a targeted preprocessing pipeline comprising Synthetic Minority Oversampling Technique, Tomek Links, Principal Component Analysis, and Robust Scaling to enhance data quality and model robustness. While deep learning approaches have recently gained traction in landslide susceptibility modeling, their efficacy typically lags that of advanced tree-based algorithms, largely due to difficulties in managing complex, high-dimensional tabular geospatial data. This research conducts a comprehensive comparative analysis of five state-of-the-art models: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM), and Attentive Tabular Neural Network (TabNet). Among the evaluated models, LightGBM achieved the highest area under the curve (AUC) of 0.980, indicating excellent ability to distinguish landslide-prone areas. However, when considering a combination of key metrics including accuracy, recall, F1-score, and precision, XGBoost demonstrated the best overall performance, while RF delivered consistently strong results across multiple criteria. While TabNet AUC (0.942) was slightly lower, its competitive recall and precision suggest it remains a promising deep learning alternative for modeling tabular geospatial data when paired with thorough preprocessing. These results highlight the effectiveness of advanced tree-based algorithms in landslide susceptibility prediction while illustrating the potential of TabNet as a complementary approach.

Tree-based and deep learning models for landslide susceptibility assessment with imbalanced data

Key Points

Abstract

Cite This Study