Predicting Cumulative Land Subsidence and Its Spatiotemporal Relationship Using Machine Learning

Land subsidence has long been a critical environmental hazard along the southwestern coast of Taiwan, with Yunlin County being one of the most severely affected areas. In this study, Long Short-Term Memory (LSTM) neural networks are employed to develop predictive models for land subsidence. Cumulative land subsidence, groundwater-level variations, and lithological layering are considered as input features to investigate the predictive performance of the models from both temporal and spatial perspectives.

As long-term groundwater monitoring data often suffer from missing values, this study further introduces a Cue Wasserstein GAN with Gradient Penalty (CWGAIN-GP) to impute missing groundwater-level data, thereby improving the stability and completeness of subsequent prediction models. Artificial masking experiments, including continuous missing periods ranging from one month to one year and random removal of 10%–50% of the data. The results show that the average Nash–Sutcliffe efficiency (NSE) achieved by the imputation model reaches 0.897.

For temporal prediction, the land subsidence model is trained using different training lengths (one year and seven years) and variable combinations to forecast cumulative land subsidence over the following one to two years. The most recent six months of observations are used as input to predict the monthly land subsidence increment. The results indicate that longer training periods and more comprehensive input variables lead to improved model performance. The coefficient of determination (R²) for the first prediction year reaches 0.945, while for the second year—under conditions of three consecutive months of missing data—the R² remains as high as 0.923.

For spatial prediction, a multi-station training and single-station validation strategy is adopted. When predicting a target station, the three nearest neighboring stations are selected, and their observations from the most recent three months are used as inputs to predict the monthly land subsidence increment at the target station. This increment is then combined with the known cumulative subsidence from the previous month to estimate the current cumulative subsidence. The results show that the average R² for single-month predictions reaches 0.966. Even when cumulative subsidence is estimated iteratively by adding predicted monthly increments over six consecutive months, the average R² remains around 0.90, demonstrating strong spatial generalization capability of the proposed model.

Fig.1 Monthly vertical profiles of cumulative land subsidence at different depths for the Huwei (MW_HWES) station in 2021.

Overall, this study demonstrates that cumulative land subsidence can be effectively predicted by integrating temporally and spatially informed LSTM models with vertically stratified hydrogeological information. Although cumulative subsidence is used as the primary prediction target, the inclusion of groundwater-level variations and lithological layering enables the model to capture the vertical characteristics of aquifer systems and their influence on subsidence processes. The results highlight the importance of incorporating stratified subsurface information when modeling land subsidence and provide a robust framework for spatiotemporal subsidence prediction under realistic data availability constraints.