Near-Decadal Land Subsidence Susceptibility and Trends Using Physics-Informed LSTM

Land subsidence poses growing risks to urban infrastructure, water resources, and long-term resilience, requiring assessment frameworks that link present-day observations with planning-relevant forecasts. This study develops an integrated approach for land subsidence susceptibility mapping and trend forecasting over multi-year horizons. The analysis uses SBAS-InSAR deformation time series derived from Sentinel-1 observations from 2017 to 2025 to characterize subsidence patterns across East Baton Rouge Parish, Louisiana. Subsidence susceptibility is modeled using an ensemble machine-learning framework that combines Extra Trees and Random Forest regressors and incorporates geological, topographic, hydrological, land use, infrastructure, and climatic conditioning factors. The susceptibility results highlight the dominant influence of land use, elevation, proximity to faults and rivers, and terrain-hydrology interactions on subsidence patterns. To extend assessment beyond observation periods, a physics-informed long short-term memory (LSTM) ensemble is introduced for forecasting. The model integrates data-driven learning with physically motivated constraints to ensure stable and realistic deformation trajectories. The forecasts preserve observed spatial patterns while exhibiting physically consistent temporal evolution and quantified uncertainty. The results demonstrate that combining InSAR observations with physics-informed deep learning enables robust, planning-scale subsidence assessment and forecasting. The proposed framework is transferable to other urban settings where long-term subsidence poses increasing societal risk.