PEGNet: A Peridynamics-Inspired and Emergent-Feature-Conditioned Spatio-Temporal Graph Neural Network for Land Subsidence Modeling

Land subsidence prediction remains challenging. Conventional grid-based or sequence-only neural networks struggle to represent these spatial dependencies and often lack structured mechanisms for incorporating region-level deformation priors and local physical consistency. This study develops PEGNet, a Peridynamics-inspired and emergent-feature-conditioned spatio-temporal graph neural network for land subsidence prediction. PEGNet uses a per-node GRU for temporal encoding and an EF-conditioned GAT for spatial aggregation. At the node level, historical InSAR deformation and auxiliary hydroclimatic observations are augmented with a PD-derived state and an EF neighborhood context, allowing a per-node GRU to learn temporal evolution and produce a base estimate of future incremental deformation. At the edge level, a fixed spatial graph supports EF-conditioned graph attention, where edge representations combine spatial distance, PD-derived bond-strain information, and the EF same/cross-region gate to generate a spatial residual prediction. At the training and prediction level, the temporal and spatial branches are fused through a learned gate, cumulative deformation is reconstructed from the last observed value, and training combines a data-fitting loss with a PD-inspired local consistency term weighted by distance and EF relations. EF therefore conditions both the network and the coupled regularizer, although it introduces no separate loss term, and the PD component remains a local consistency mechanism rather than a complete Peridynamic solver. PEGNet is evaluated using 60 months of PS-InSAR observations in Tongzhou District, Beijing, four overlapping purged test windows covering nine unique target months and five random seeds. It achieves an incremental-deformation RMSE of 2.153 ± 0.007 mm and a reconstructed cumulative-deformation RMSE of 2.631 ± 0.072 mm. These results outperform the deterministic baselines and remain comparable to the unconstrained GRU-GAT model. Mechanism diagnostics indicate that EF mainly redistributes graph attention, whereas the PD components improve local consistency without producing a substantial global accuracy gain. Overall, the comparable aggregate performance and improved local consistency support the use of PEGNet for conditional subsidence monitoring and scenario analysis.