Observation-Integrity- and Geometry-Aware SimVP-RD for Irregular PS-InSAR Deformation Forecasting of Earth-Rockfill Dams

PS-InSAR provides spatially distributed line-of-sight (LOS) deformation measurements, creating an opportunity to forecast dam-surface deformation as a spatial field rather than at isolated monitoring points. However, irregular acquisition intervals, sparse observations, and anisotropic radar pixels complicate multi-step forecasting. We develop SimVP-RD, an observation-integrity- and geometry-aware framework that predicts reference-interval increments, scales them by the actual elapsed time, and integrates them from the latest observed field. The framework combines fixed-support masking, a reconstruction-dynamics objective, native-grid anisotropic convolution, and recursive rollout training. Using 94 real acquisitions of Lianghekou Dam, SimVP-RD achieves an MAE of 3.355 ± 0.024 mm and an RMSE of 4.534 ± 0.034 mm under a matched observed-only 8→8 recursive protocol, outperforming ConvLSTM, PredRNN, and TAU. The corrected Stage-A ablation further shows that the reconstruction-dynamics objective reduces MAE from 3.833 ± 0.111 to 3.549 ± 0.086 mm. On the native PS grid, the 3 × 7 convolution retains all 2170 valid locations and provides a more balanced physical receptive field than the conventional 3 × 3 kernel. Spatial and interval-stratified diagnostics further characterize how forecast errors vary across prediction leads and spatial contexts. These results establish SimVP-RD as a support-preserving approach to multi-step LOS deformation-field forecasting under irregular satellite sampling.