Machine Learning (ML) Approaches in Landslide Prediction: A Systematic Review and Implications for Environmental Risk Management and Science Education

Abstract

This study systematically reviewed the application of machine learning (ML) models in landslide prediction, synthesizing advances, limitations, and emerging directions. Using the PRISMA framework, eight studies published between 2019 and 2024 were selected from major databases following rigorous inclusion criteria. Results revealed that ML models varied from standalone to ensemble, hybrid, and optimized approaches, with ensemble and hybrid frameworks often demonstrating superior reliability across diverse geospatial contexts. Performance metrics such as training and validation AUC, accuracy, and recall indicated generally strong predictive capacity, though inconsistencies in reporting limited cross-study comparability. Conditioning factors and preprocessing strategies were highly variable, reflecting both regional specificity and methodological divergence, while validation techniques ranged from random splits to cross-validation. Key challenges included imbalanced datasets, a lack of standardized metrics, and limited integration of geophysical and environmental factors, which constrained model transferability and generalizability. Emerging technologies such as explainable AI, transfer learning, advanced GIS integration, and multi-temporal remote sensing were identified as promising avenues to address these gaps. Overall, the findings highlight the capacity of ML to advance landslide susceptibility mapping while underscoring the need for methodological standardization and cross-regional datasets, with broader implications for disaster risk reduction, sustainability, and science education.