AI-Based Displacement Forecasting for Real-Time Landslide Risk Assessment in the Danube Region

Abstract

This paper presents a dual-layered AI framework for real-time landslide risk assessment developed under the GeoNetSee project within the Interreg Danube Region Programme. The first layer employs a fuzzy logic model, inspired by the Slovenian MASPREM system, which integrates Landslide Susceptibility Maps (LSS) with high-resolution precipitation forecasts from the Open-Meteo API to generate a Predicted Landslide Hazard (PLSH) score on a 0–5 scale, updated every 6–12 hours. The second layer focuses on real-time ground displacement detection by fusing low-cost dual-frequency GNSS receivers with MEMS accelerometers and applying machine learning regression algorithms, including Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Convolutional Neural Network – Long Short-Term Memory (CNN-LSTM). Validation at the Curine Njive pilot site in Bosnia and Herzegovina demonstrated that SVR achieved a Mean Absolute Error (MAE) as low as 0.0018 m across four GNSS nodes, while the CNN-LSTM achieved the highest overall predictive accuracy across all evaluated nodes with a MAPE of only 3.99% for Node 41. The fuzzy model successfully identified critical hazard levels during intense rainfall events. Results confirm the viability of integrating low-cost sensor networks with AI-driven analytics for transboundary geohazard early warning systems.