Forecasting Deep-Seated Landslide Displacements Using Machine Learning and Automated Monitoring Data

Abstract

Deep-seated landslides are slope failures where the primary sliding surface extends beyond the soil mantle into weathered or intact bedrock, involving large volumes and slow displacement rates. Forecasting their displacements is challenging due to complex hydrological, geotechnical, and climatic interactions. This study develops a machine learning framework for short-term displacement forecasting at 6 m depth using hourly automated monitoring data from a tectonically active landslide in Lam Dong Province, Vietnam. Three models, Gradient Boosting (GB), Support Vector Regression (SVR), and Multilayer Perceptron Regression (MLPR), were trained on 16 hydrometeorological and geotechnical variables from January 2021 to November 2024. Among them, MLPR achieved the best performance (R² = 0.920; MAE = 0.036 mm; MAPE = 1.57%), surpassing GB (R² = 0.891) and SVR (R² = 0.873). Residual and partial dependence analyses confirmed the robustness and interpretability of MLPR, identifying precipitation, pore water pressure, and volumetric water content as dominant predictors. The results demonstrate that integrating multi-sensor real-time data with ML improves displacement forecasting accuracy and timeliness, enhancing early-warning and mitigation strategies. While this model is site-specific, it provides a scalable foundation for hybrid ML–physics approaches and multi-site ensemble learning, enabling generalization across diverse geomorphic conditions.