Research on dam deformation monitoring technology based on time-series InSAR and BeiDou

Abstract

Time-series synthetic aperture radar interferometry provides high-density deformation observations but only resolves displacement along the radar line of sight. A monitoring method combining small baseline subset InSAR, BeiDou three-dimensional measurements, residual-variance-based two-stage fusion, and a long short-term memory network is developed. In the first stage, the BeiDou components constrain the decomposition of the InSAR line-of-sight displacement; in the second stage, the fusion weights are updated from residual variances within a rolling window. InfluxDB organizes the observations by timestamp, station, data source, and deformation component to support windowed alignment and repeated residual retrieval. Station-block bootstrap resampling is introduced to quantify 95% confidence intervals for the root mean square error and mean absolute error. Tests using the Xiaolangdi Dam dataset from 2017 to 2023 show that the fusion method maintains three-dimensional deformation root-mean-square error (RMSE) below 3 mm and mean absolute error (MAE) below 2.5 mm. When the prediction horizon is extended to 300 days, the coefficient of determination remains 0.817. The comparison with physics-informed and Transformer-based predictors indicates that the improvement mainly results from the spatially enriched three-dimensional input rather than a modified LSTM gate.