InSAR and GAT-LSTM integration for dam displacement prediction: Lessons from the Oldman River Dam, Canada

The precise prediction of dam deformation is essential for ensuring infrastructure safety and mitigating geohazards, particularly in regions characterized by limited monitoring studies. This research concentrates on the Oldman River Dam in Alberta, Canada, where Interferometric Synthetic Aperture Radar (InSAR)-based deformation monitoring and prediction remain inadequately explored. A novel framework that integrates a Graph Attention Network with Long Short-Term Memory (GAT-LSTM) has been developed to address the limitations of existing methods, which neglect spatial dependencies among InSAR-derived points and the increased model complexity stemming from point clustering or InSAR time series decomposition. Sentinel-1 data from three passes were processed utilizing a full-resolution InSAR technique, resulting in semi-vertical deformation velocities that demonstrated consistent subsidence along the dam crest, with rates fluctuating from 5.08 to 6.23 mm/yr. A robust correlation between deformation and reservoir water levels was noted, with accelerated crest deformation during the 2017–2019 drawdown period and a potential risk identified due to a significant decline in water levels projected for 2023–2024. The GAT-LSTM model, which captures both spatial and temporal dynamics, outperformed the standard LSTM, achieving 83.64% accurate points compared to 76.90% for the LSTM in short-term forecasting, exhibiting notable reliability along the crest. The peak performance was observed on September 9, 2021, with a Root Mean Square Error of 0.30 ± 0.013 mm and a Mean Absolute Error of 0.22 ± 0.012 mm. The proposed framework would enhance dam safety monitoring by providing actionable short-term predictions, demonstrating potential transferability to other slow-moving infrastructure. • Developed a GAT-LSTM framework for dam deformation prediction using time series InSAR data. • Demonstrated strong correlation between water level and dam crest subsidence. • GAT–LSTM outperformed the conventional LSTM, achieving a best RMSE of 0.30 mm and MAE of 0.22 mm. • Provides a tool for early warning in dam safety management.