High resolution estimates of water table depth across Brazil: insights into model behavior and data challenges

Abstract

Groundwater plays an important role in sustaining Brazil’s ecosystems and growing commercial agriculture. However, national-scale understanding remains limited due to sparse monitoring (<0.02 wells per km2). This study develops the first high-resolution (90 m) national-scale map of water table depth (WTD) over Brazil using a random forest model and characterizes model behavior and data challenges. Trained on available groundwater observations and datasets on climatology, topography, hydrogeology, and soil characteristics, the model achieves moderate predictive skill (test r = 0.61; RMSE = 19 m; MAE = 10.41 m) and reproduces known hydrologic patterns in major basins. We provide an estimate of model uncertainty for each prediction at the grid cell level, showing deeper water tables carry greater uncertainty. Permutation feature importance and SHapley Additive exPlanations show that height above nearest drainage, hydraulic conductivity and elevation are key in WTD predictions and that the model captures important physical relationships (e.g. predicted WTD increases with elevation). Feature analysis also allows us to identify that the hydraulic conductivity dataset used—derived from a global product—is not representative of hydrogeologic realities in Brazil. In addition, we show how model performance scales with the amount of training data, but where these observations are matters more. We highlight the need for more groundwater monitoring wells, particularly in sparsely monitored regions. This national-scale, high-resolution map of WTD advances our understanding of the spatial variability in Brazil’s groundwater resources. The work also offers insights into improving machine learning model interpretability and highlights the potential of data-driven models to help improve global WTD estimates, particularly in other data-limited geographies.