TCN-GCN: a spatiotemporal neural network for multistep InSAR deformation prediction in geologically diverse regions

Abstract

Accurate multi-step prediction of surface deformation monitored by Interferometric Synthetic Aperture Radar (InSAR) is critical for proactive disaster early warning in complex engineering environments. Existing approaches either neglect spatial dependencies among monitoring points or rely on recurrent architectures that suffer from gradient vanishing over extended prediction horizons. This paper proposes TCN-GCN, a spatiotemporal neural network that integrates a Temporal Convolutional Network (TCN) for long-range temporal feature extraction with a Graph Convolutional Network (GCN) for encoding spatial correlations among InSAR coherent scatterer points. Monitoring points are modeled as graph nodes with edges weighted by a hybrid function of spatial proximity and deformation correlation. The proposed framework is evaluated on two geologically distinct InSAR datasets: reservoir-induced deformation in the Three Gorges–Gezhouba area, Hubei Province, and coastal road settlement in Xiapu County, Fujian Province. TCN-GCN achieves RMSE reductions of 33.0 and 27.2 over the strongest baseline (LSTM) on the two datasets respectively, and maintains stable prediction accuracy across forecasting horizons up to 12 time steps.