Application of Machine Learning in Reconstructing Continuous Daily Water Table Depth from Sparse Data and Wetland Hydrology Assessment in the Forested Wetland Systems

Abstract

Water table depth (WTD) is a key control on wetland ecology, hydrology, and forest productivity in coastal plain landscapes, yet continuous groundwater records are rarely available. This limitation makes it difficult to quantify climate controls on groundwater response and derive wetland hydrology indicators. This study addressed these gaps in the Santee Experimental Forest, South Carolina, using 13 wells across three watersheds. The test data were selected within each well using a structured holdout that retained early, middle, and late observations for evaluation. Two objectives were pursued: examining climatic controls on daily WTD change and reconstructing daily WTD series from sparse observations using machine learning models. For prediction with climatic predictors and lagged WTD, all evaluated models performed strongly (NSE = 0.948–0.981; RMSE = 8.53–14.12 cm), with Linear Regression as the baseline and XGBoost performing best overall. For reconstruction without lagged WTD, performance declined more noticeably, and all models outperformed Linear Regression, with the ensemble tree models having the edge over the others. XGBoost showed the strongest reconstruction performance among the tested models and was therefore used for the downstream trend and wetland-hydrology analyses, with NSE of 0.763, KGE of 0.849, RMSE of 29.86 cm, NRMSE of 0.487, MAE of 19.24 cm, and KS of 0.072. Sensitivity analysis showed that the climate-composite predictor set gave the strongest reconstruction skill, while soil moisture only with day of year also gave good results. These results show that machine-learning reconstruction can extend sparse monitoring and support hydrologic interpretation and wetland hydrology evaluation.