InSAR Deformation Prediction for Land Reclamation Integrating Consolidation Model and Transformer

Objectives: Large-scale coastal land reclamation often triggers prolonged and uneven subsidence in soft-soil areas, making reliable deformation prediction essential for infrastructure safety. Existing methods face inherent limitations: physics-based consolidation models capture primary trends but generalize poorly to complex nonlinear signals; statistical machine learning approaches such as SVR and RF lack physical constraints and extrapolate unreliably; and deep learning architectures like LSTM, Transformer and Informer rarely incorporate domain-specific physical knowledge, compromising interpretability and physical consistency. Methods: A hybrid framework is developed that tightly couples a consolidation model with a Transformer encoder. The InSAR deformation time series is first decomposed into a physically governed consolidation trend and a residual component. A consolidation model is fitted to extract the long-term subsidence trend, while the residual—containing seasonal fluctuations, secondary compression, and anthropogenic disturbances—is modeled by a Transformer encoder. Multi-head self-attention captures long-range temporal dependencies within the residual, and an attention pooling mechanism adaptively emphasizes the most critical time steps for forecasting. The final prediction is obtained by summing the extrapolated physical trend and the predicted residual. Validation is conducted at three reclaimed coastal sites in Fujian Province—Funing Bay (Ningde), Fuzhou Changle International Airport, and Shuangyu Island (Zhangzhou)—using 662 Sentinel-1A images processed by time-series InSAR. Results: The hybrid model achieves the highest test-set accuracy across all sites. Compared with SVR, RF, a standalone consolidation model, a standard Transformer, LSTM, and Informer, RMSE is reduced by 87.3%, 83.7%, 54.7%, 20.4%, 24.7%, and 10.0%, respectively. The drastic improvement over the pure consolidation model confirms the necessity of capturing residual dynamics, and the clear gains over deep learning baselines demonstrate the value of physical priors. Conclusions: Integrating a consolidation model with Transformer-based deep learning markedly enhances InSAR deformation forecasting accuracy and physical credibility. The proposed approach provides an effective tool for subsidence monitoring, early warning, and risk mitigation in coastal reclamation zones.