Hybrid deep learning approach for predicting long-term deformation of landslides

Actively moving landslides threaten life, property, and infrastructure. While monitoring movement velocity is essential for understanding landslide processes, predicting deformation magnitude is a significant step toward mitigation. This study develops a hybrid deep learning model integrating a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and attention mechanisms to predict landslide-related ground deformation from spatial and temporal data. The model was developed for areas of Central Italy heavily affected by the 2016–2017 seismic events, using a three-way experimental protocol. Temporal data comprise Persistent Scatterer (PS) observations for 2018–2022, whereas spatial data represent landslide predisposing factors. The CNN extracts spatial features associated with landslide susceptibility, while the LSTM captures temporal dependencies in cumulative deformation. Attention mechanisms further refine the model by weighting critical regions and time steps. On an independent hold-out test set, not used for hyperparameter tuning or model selection, the model achieved low prediction errors for both ascending and descending geometries, and the close agreement between training, validation, and test results indicates limited overfitting. Overall, the proposed CNN-LSTM-attention framework provides a transferable, InSAR-driven basis for anticipating where and when landslide deformation is likely to intensify, supporting hazard zonation, monitoring prioritisation, and early-warning strategies in seismically active regions.