Spatiotemporal Dam Deformation Monitoring and Prediction using InSAR and Deep Learning

Monitoring dam deformation is critical for mitigating geohazards and ensuring the safety of both water-retaining and tailings dam infrastructure. Conventional in situ monitoring techniques provide accurate point-based measurements but are spatially sparse and do not cover the whole dam structure. Interferometric Synthetic Aperture Radar (InSAR) complements the traditional dam monitoring techniques, providing observations of surface deformation over the entire structure. In this study, we present an InSAR-based monitoring and prediction framework applied to two dams: the Oldman River dam in Alberta, Canada, and the Córrego do Feijão Tailings Dam I in Minas Gerais, Brazil. Sentinel-1 SAR data were processed using an InSAR time-series technique to derive detailed deformation patterns over the two dam sites. At the Oldman River Dam, semi-vertical deformation velocities revealed consistent subsidence along the dam crest, with rates ranging from 5.08 to 6.23 mm/yr. The observed deformation exhibited a temporal relationship with fluctuations in reservoir water levels, including accelerated crest deformation during the drawdown period. In contrast, pre-failure deformation analysis of the Córrego do Feijão Tailings Dam I revealed pronounced deformation behind the crest, with line-of-sight velocities reaching up to −69 mm/yr prior to the catastrophic failure in January 2019.  To address the limitations of conventional time-series prediction approaches, particularly their inability to account for spatial dependencies among InSAR measurement points, a graph-based deep learning architecture that explicitly models spatial relationships was introduced. Specifically, spatiotemporal Graph Attention Network (GAT)–based recurrent models, namely GAT-Long Short-Term Memory (GAT-LSTM) and GAT-Gated Recurrent Unit (GAT-GRU), were proposed to jointly capture spatial dependencies and temporal dynamics in InSAR deformation data. The proposed models outperformed equivalent non-graph recurrent neural network baselines (i.e., LSTM and GRU) in deformation forecasting. Overall, the results demonstrated the robustness and transferability of InSAR-driven, graph-based predictive frameworks for diverse dam environments. The proposed approach provides a scalable pathway for deformation monitoring and early warning systems, enabling proactive risk management for critical dam infrastructure worldwide.