Satellite data and physics-constrained machine learning for estimating effective precipitation in the Western United States and application for monitoring groundwater irrigation

Effective precipitation, the portion of evapotranspiration derived from precipitation, is an important part of the agricultural water balance and affects the amount of water required for irrigation. Due to hydrologic complexity, effective precipitation is challenging to quantify and validate using existing empirical and process-based methods. Moreover, there is no readily available high-resolution effective precipitation dataset for the United States (US), despite its importance in determining consumptive use of irrigation water. Here, we developed a framework that incorporates multiple hydrologic states and fluxes within a machine learning approach that accurately predicts effective precipitation for irrigated croplands of the Western US at ∼2 km spatial resolution and monthly scale from 2000 to 2020. We analyzed the factors influencing effective precipitation to understand its dynamics in irrigated landscapes. To further assess effective precipitation estimates, we estimated groundwater pumping for irrigation in seven basins of the Western US with a water balance model incorporating model-generated effective precipitation. A comparison of our estimated pumping volumes with in-situ records indicates good skill, with R2 of 0.78 and PBIAS of –15 %. Though challenges remain in predicting and assessing effective precipitation, the satisfactory performance of our approach illustrate the application and potential of integrating satellite data and machine learning with a physically-based water balance to estimate key water fluxes. The effective precipitation dataset developed in this study has the potential to be used with satellite-based actual evapotranspiration data for estimating consumptive use of irrigation water at large spatio-temporal scales and enable the best available science-informed water management decisions.