Evaluating model robustness in landslide susceptibility mapping using a unified data-consistent framework in northern Thailand

Abstract

Abstract Landslide susceptibility mapping (LSM) is a critical tool for hazard mitigation in mountainous regions. However, the reliability of existing models remains uncertain due to inconsistencies in data quality, sampling strategies, and validation approaches. Although machine learning (ML) and deep learning (DL) models often report high predictive accuracy, their performance may not be robust or transferable, particularly in data-constrained tropical environments. To address this limitation, this study develops a unified, data-consistent evaluation framework to compare statistical, ML, and DL approaches for landslide susceptibility mapping in the Luang Prabang Range, northern Thailand. Five models, including Frequency Ratio (FR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM), were applied using the same landslide inventory, conditioning factors, and validation strategy to ensure a fair comparison. The results indicate that RF achieved the most stable and reliable performance (AUC = 0.942), followed by XGBoost (0.930), CNN (0.914), LSTM (0.902), and FR (0.865). While DL models demonstrated strong predictive capability, their performance was more sensitive to data limitations and model configuration. In contrast, RF provided a better balance between accuracy, robustness, and data efficiency. The findings demonstrate that differences in model performance are influenced not only by algorithm selection but also by data structure and parameter settings. The main contribution of the study is the implementation of a unified evaluation framework that enables more reliable assessment of model robustness, uncertainty, and interpretability. This study provides practical guidance for model selection in landslide susceptibility mapping and highlights the importance of data consistency and model transparency, particularly in tropical mountainous regions where data limitations are common.