A Novel ML-Based Methodology for Estimating Water Table Depth Anomalies at the European Continent Scale

Effective and efficient groundwater monitoring at the continental scale is still a challenge, mainly due to the scarcity of water table depth (wtd) observations. In this study, we proposed a novel methodology based on advanced machine learning techniques, Long Short-Term Memory (LSTM) networks and transfer learning, for estimating monthly wtd anomalies (wtda) over Europe with monthly precipitation and soil moisture anomalies (pra and a) as input. In the methodology, the LSTM networks were trained on simulation results and then, without additional training, utilized to estimate wtda based on pra and a from observational datasets, thereby transferring the simulated input-output relationship to the observation-based estimation. The obtained estimates were evaluated based on in-situ wtd measurements at 2,604 wells distributed over different European regions, achieving R from 0.39 to 0.79 and RMSE from 0.37 to 1.1 for regional averaged values. This constitutes of a >0.10 increase in R and a >0.13 decrease in RMSE compared to the simulation results used for training. In addition, compared with the LSTM networks directly trained on observations, the proposed methodology showed slightly worse test performance at the individual pixel level, lending confidence to applications in areas without wtd observations. The study provides a validated methodology for producing reliable wtda estimates over the European domain in the absence of observations, which can be used for data reconstruction and online groundwater monitoring useful in European groundwater management.