Advancing Large-Scale Hyper-Resolution Groundwater Modeling Using a Machine Learning-Based Downscaling Tool

Groundwater is becoming more important in sustainable water management, particularly in the context of climate change and intensive human interventions. Given that groundwater varies in space and time, it is important to predict both its dynamic processes and static patterns. However, lack of reliable groundwater data restricts the development of large-scale groundwater monitoring systems linking observations with modeling at spatial scales relevant for local decision making. In this study, we leverage existing physically-based modeling data and water table depth observations in the Contiguous United States (CONUS) and develop a machine learning-based downscaling tool to downscale 1-km modeling data to 1arcsec (~ 30 m). The modeling data were generated daily for the water year 2003 using ParFlow, a three-dimensional integrated hydrologic model. In addition, we input a range of meteorological, topographic, geological, and land use data, including daily precipitation and temperature, elevation, hydraulic conductivity, mean soil and clay contents, and land cover types. Based on tree-based machine learning models running on GPUs, the downscaling tool outputs a 1 arcsec water table depth map for the CONUS daily in a relatively short time. The resulting hyper-resolution water table depth map incorporates groundwater pumping and uncertainty, significantly advancing our understanding of groundwater dynamics across various scales, from the continental to small scales relevant to local decision-making. We also obtain the importance of input variables based on the results of the machine learning models, which is helpful for the future development of the groundwater monitoring system over the CONUS. While the downscaling tool is developed for the CONUS, it can be adapted to other regions with similar hydrogeological settings and substantial modeling data.