A physics-informed neural network approach to predicting land subsidence-rebound in Dezhou City under different climate scenarios

Dezhou, China, is one of the typical areas of land subsidence in the North China Plain. This study focuses on the changes in subsidence patterns following groundwater level (GWL) recovery in Dezhou City, developing a dualistic water cycle framework that integrates both climate and human factors. Two Physics-Informed Neural Network (PINN) models are constructed to simulate: (1) the relationship between precipitation, evapotranspiration, groundwater (GW) extraction, and GWLs (2) the coupling between GWLs and land subsidence. Trained with meteorological, hydrogeological, and Interferometric Synthetic Aperture Radar (InSAR) deformation data, the models use Shared Socioeconomic Pathway–Representative Concentration Pathway (SSP-RCP) scenario data and simulated GW extraction data to predict future GWLs and subsidence under different scenarios. Shallow GWLs are highly sensitive to climate change, showing significant seasonal fluctuations under the SSP5-RCP8.5 scenario, with a maximum amplitude of 2.79 m. In contrast, deep GWLs have a slower response, though long-term trends gradually emerge under the SSP5-RCP8.5 scenario, up to 0.872 m/yr. Groundwater extraction directly drives GWL decline, suppressing seasonal fluctuations and extending the response time to precipitation, with a maximum lag of 8 months. Precipitation indirectly affects subsidence through the multi-aquifer system, with subsidence-rebound variations mainly influenced by groundwater extraction and GWL fluctuations. Overall, climate change affects subsidence fluctuations, while groundwater extraction remains the primary factor for long-term subsidence trends. • Constructed a regional dualistic water cycle framework linking climate and human impact. • Applied physics-constrained WB-RNN model for groundwater level inversion. • Proposed consolidation theory-constrained deep learning model for subsidence. • Predicted Dezhou land subsidence under various SSP-RCP and extraction scenarios.