LGTM: A Time Series Model for Large-Scale Geospatial Land Subsidence Prediction

Land subsidence threatens infrastructure, ecosystems and coastal communities, yet predicting its spatially heterogeneous evolution from long-term interferometric synthetic aperture radar (InSAR) imagery remains challenging. Here, we propose the Large-scale Geospatial Time-series Model (LGTM), a spatiotemporal fusion network that combines a dual-path Informer-p temporal encoder with a simplified U-Net spatial decoder to model long-range temporal dependencies and multiscale spatial patterns. Using 136 cumulative SBAS-InSAR deformation images acquired from 10 January 2018 to 1 June 2023 in the Yellow River Delta, China, we evaluated three approximately 943.7 km² subregions representing circular subsidence funnels, irregular patches and radially contiguous subsidence clusters. Under identical data partitions, normalization procedures, training protocols and evaluation masks, LGTM was compared with ConvLSTM, TimesNet and BiLSTM, while ablation experiments examined Informer, Informer-p and Informer+U-Net. LGTM achieved the highest structural similarity index (SSIM) in all three regions, whereas histogram intersection over union (IoU) and normalized cross-correlation (NCC) were highest or tied for highest. The largest relative reduction in root mean square error (RMSE) and increases in SSIM, IoU and NCC were 36.90%, 7.57%, 3.10% and 6.39%, respectively. Although LGTM did not achieve the lowest RMSE in every region, it consistently outperformed Informer+U-Net across all four metrics in the ablation experiments. Informer-p alone did not consistently outperform the original Informer, indicating that the observed gains were associated primarily with the joint temporal-spatial fusion. These results demonstrate that LGTM provides a balanced approach to numerical prediction and preservation of complex subsidence structures in regional InSAR time-series forecasting.