Image: Water: "Groundwater Level Prediction Using Machine Learning and Geostatistical Interpolation Models" Authors: Fabian J. Zowam1 and Dr. Adam M. Milewski1 1-Water Resources & Remote Sensing Laboratory (WRRS), Department of Geology, University of Georgia, Athens, GA 30602, USA Correspondence: Dr. Adam Milewski NOTE: Exciting new developments in groundwater management! Doctoral candidate, Fabian Zowam, and Professor of Hydrogeology, Dr. Adam Milewski, publish a new study on Arizona's first statewide prediction of groundwater level anomalies (GWLA), which is crucial for addressing climate change's impacts on surface water. Using machine learning, we forecasted monthly GWLA from January 2010 to December 2019 across multiple aquifers. Our multi-model approach showed good predictive accuracy. More importantly, our remote sensing-based model can be adapted for other regions facing similar challenges, ensuring that even data-sparse areas aren’t overlooked. This work highlights the importance of innovative strategies in resource management as we navigate the realities of a changing climate. ABSTRACT: Given the vulnerability of surface water to the direct impacts of climate change, the accurate prediction of groundwater levels has become increasingly important, particularly for dry regions, offering significant resource management benefits. This study presents the first statewide groundwater level anomaly (GWLA) prediction for Arizona across its two distinct aquifer types—unconsolidated sand and gravel aquifers and rock aquifers. Machine learning (ML) models were combined with empirical Bayesian kriging (EBK) geostatistical interpolation models to predict monthly GWLAs between January 2010 and December 2019. Model evaluations were based on the Nash–Sutcliffe efficiency (NSE) and coefficient of determination (R2) metrics. With average NSE/R2 values of 0.62/0.63 and 0.72/0.76 during the validation and test phases, respectively, our multi-model approach demonstrated satisfactory performance, and the predictive accuracy was much higher for the unconsolidated sand and gravel aquifers. By employing a remote sensing-based approach, our proposed model design can be replicated for similar climates globally, and hydrologically data-sparse and remote areas of the world are not left out. Keywords: groundwater level; machine learning; empirical Bayesian kriging; remote sensing Type of News/Audience: Department News Research Areas: Environmental Geosciences
