LSTM-based projection of groundwater-induced land subsidence under CMIP6 climate scenarios: A case study of Yunlin, Taiwan

Abstract

Land subsidence is usually driven by excessive groundwater extraction and further intensified by climate variability. Monitoring land subsidence and projecting its future trends are foundational for sustainable groundwater and subsidence management. This study develops a deep learning-based framework to project future land subsidence in Yunlin County, a major subsidence hotspot within the Choushui River Alluvial Fan in Central Taiwan, under CMIP6 climate change scenarios. A Long Short-Term Memory (LSTM) model is employed to estimate future subsidence rates from 2022 to 2036 by using deep learning-driven groundwater level projections under SSP scenarios. The model is trained using historical groundwater level and multi-level compaction well data from 2015 to 2021. The proposed model accurately estimates land subsidence associated with groundwater level variations, achieving low RMSE (<0.4 cm), NRMSE (<0.1), MAE (<0.3 cm) and high R² (∼0.8) as well as correlation (R) (∼0.9) values. Projections reveal pronounced spatial heterogeneity in future subsidence, with coastal and central regions showing heightened vulnerability. Under SSP245 and SSP585 scenarios, subsidence rates could accelerate historical averages by 1.5-2 cm/year in agricultural-intensive regions, driven by elevated temperatures, declining precipitation causing increased pumping demand, thereby, intensifying groundwater decline. Pronounced effects are further observed for the higher emission SSP585 scenario. Additionally, artificial subsidence mitigation scenarios simulating 10-20 % reductions in groundwater extraction demonstrate the model's effectiveness in assessing intervention strategies, showing a 16-50 % drop in simulated compaction. This research demonstrates the compounding impacts of climate change on groundwater-induced subsidence and emphasizes the effectiveness of deep learning in spatio-temporal forecasting, offering critical insights for adaptive groundwater management and subsidence risk mitigation in regions facing coupled climatic and anthropogenic stressors.