PREDICTION OF GROUNDWATER LEVEL ANOMALY BASED ON GRACE AND WATER-BUDGET VARIABLES USING MACHINE LEARNING IN THE GLACIAL AQUIFER SYSTEM

Nearly 41-million people rely on withdrawals from the glacial aquifer system in the US. As a result, these aquifer systems are vulnerable to climate and anthropogenic impacts. Thus, monitoring the terrestrial water cycle is essential to better manage the water resources in these systems. The Gravity Recovery and Climate Experiment (GRACE) satellite provided unprecedented information regarding the terrestrial water cycle, however, it is difficult to utilize GRACE for local scale applications due to its coarse spatial resolution. This study proposes a downscaling approach to predict monthly groundwater level Anomaly (GWLA) at finer spatial resolution from GRACE and other water-budget variables (e.g., precipitation, streamflow, soil moisture, plant canopy surface water, and snow water). Multi-level MLMs were constructed to simulate GWLA in time and space using 32 groundwater observation wells throughout the glacial aquifer system in Illinois.

Preliminary results showed that the MLM satisfactorily predict GWLA in time and space with Nash-Sutcliffe model Efficiency (NSE) values from 0.83 to 0.98 and -1.87 to 0.99 and coefficient of correlation (r) from 0.91 to 0.99 and 0.36 to 0.99, respectively, over the entire simulation period (2002–2016). Further, using independent dataset, the models were validated with NSE from -14 to 0.89 and r from -0.27 to 0.94 for the period of Aug. 2015 – Jul. 2016. A variable importance of water-budget highly varied with watershed characteristics of each well, while GRACE TWSA remained steady. Uncertainties may arise due to outliers resulting from extreme conditions, which may not be fully captured by the MLM due to limited sample size. In addition, predictor variables used in the model may not adequately capture aquifer heterogeneity. This downscaling approach has demonstrated a potential to provide local-scale GWLA data that can be integrated into local water resources management applications.