Estimation of Monthly Water Table Depth Anomalies Based on the Integration of GRACE and ERA5-Land with Large-Scale Simulations Using Random Forest and LSTM Networks

Abstract Increasing pressure on groundwater resources, exacerbated by climate change, highlights the need to develop advanced methods for monitoring groundwater storage and levels. While numerical and physics-based models are widely used to analyze the spatial and temporal dynamics of groundwater levels, they require extensive input data and can be computationally expensive for high-resolution and large-scale simulations. In contrast, remote sensing products such as the Gravity Recovery and Climate Experiment (GRACE) provide global-scale information on total water storage anomalies. However, due to its coarse spatial resolution (0.25 $$^{\circ }$$ ), GRACE data cannot be used directly to assess groundwater conditions at local and regional scales. In order to obtain local groundwater levels that can be quickly accessed by stakeholders to monitor and define appropriate groundwater management, this study implements a methodology based on data-driven models to estimate monthly water table depth anomalies (wtda), integrating simulations from the Terrestrial Systems Modeling Platform (TSMP) with GRACE and reanalysis ERA5-Land datasets. Considering the spatial resolution of current TSMP simulations (TSMP-G2A - 0.11 degrees), we tested and compared multiple Random Forest (RF) and LSTM networks at the pixel scale over the Seine River Basin, combining different hydrological and climatological variables with GRACE as input features. For each data-driven approach, we selected the model that best represents the temporal pattern of the wtda during the test period and compared the results with the original TSMP simulation, as well as in-situ groundwater observations. The results indicate that both RF and LSTM networks can well reproduce the temporal patterns of groundwater levels across the Seine Basin obtained by the TSMP simulations, with average Pearson correlations of 0.65 and average KGE of 0.6, respectively. A comparison with multiple groundwater wells allowed us to identify the regions where the applied models are more reliable for representing wtda over the Seine River Basin. In general, the proposed models show good agreement with in-situ observations, independent of the groundwater well depth. However, we found significant differences between observed and simulated water table depths in the downstream regions of the Seine River Basin, where coastal systems and the presence of karst in the chalk might influence groundwater levels and the performance of the adopted models, respectively. The proposed methods provide end users with an extremely lightweight reconstruction and prediction tool for wtda at the pixel level, including reliability estimates, which is easy to implement in an ad hoc fashion in any evaluation and groundwater management workflow.