Answers are grounded in this paper and cite their sources.
Reclaimed coastal areas are highly susceptible to uneven ground deformation, posing significant risks to urban infrastructure. This study investigates the long-term subsidence evolution of Hengqin Island by utilizing 105 Sentinel-1A scenes acquired between September 2016 and April 2025 via persistent scatterer interferometric synthetic aperture radar. Existing prediction methods often struggle to effectively capture the complex nonlinear spatiotemporal correlations inherent in subsidence data. To address this limitation, we propose a novel spatial-temporal graph convolutional network-transformer (STGCN-Transformer) model. Unlike conventional approaches that treat spatial and temporal features in isolation, our model jointly learns coupled spatiotemporal patterns. Specifically, it utilizes STGCN to extract local spatial dependencies based on geographic proximity topology, while integrating a Transformer with a multihead self-attention mechanism to capture long-range temporal dependencies within deformation sequences. Comprehensive comparative experiments against CNN, LSTM, and standalone Transformer baselines demonstrate the superiority of the proposed framework, which achieved the highest accuracy across RMSE, R2, MAE, and sMAPE metrics. Based on the validated model, we forecast a continued subsidence trend, with a maximum cumulative settlement projected to reach −64.23 mm by January 2029. Furthermore, a four-level land subsidence risk zonation map was generated based on the predictions, providing actionable insights for urban planning and disaster mitigation in reclaimed areas. This study establishes a robust framework for high-precision InSAR deformation prediction and provides critical scientific support for disaster prevention and sustainable urban planning in complex reclaimed environments.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: