A Machine Learning Approach to Estimating Evapotranspiration and Applications for Prediction Under Future Climate Scenarios

Abstract:

Evapotranspiration (ET) constitutes the largest outgoing flux of water in the terrestrial hydrologic cycle, yet sparse in-situ observational networks have historically limited the availability of ET information for the hydrologic and atmospheric science communities. Accelerating advances in remote sensing, including ever-increasing satellite missions, computational processing power, and modelling proficiency, are fueling a revolution in the estimation of global evapotranspiration at fine spatial resolutions. In this research, we develop statistical models to approximate high-resolution annual evapotranspiration rates under future climate scenarios to better understand the impact of climate change on the hydrologic cycle. We apply the latest version of the Operational Simplified Surface Energy Balance (SSEBop) model implemented within the inter-agency cooperative OpenET project to 26 regions across the conterminous United States, encompassing almost 2.2 million square kilometers. Using precipitation, temperature, and topography information, we train and test linear regression, classification and regression trees (CART), and random forest machine learning models to approximate SSEBop evapotranspiration for each of the National Land Cover Database vegetation classifications (excluding irrigated lands). By using only two temporally variant predictors (i.e. precipitation and temperature) and three temporally invariant predictors (i.e. elevation, slope, and aspect), trained models can be applied easily across a wide range of domain sizes. We investigate the use of these statistical models to approximate ET, as well as associated water balance-derived runoff. Using precipitation and temperature estimates derived from the Global Circulation Model runs conducted under the Coupled Model Intercomparison Project Phase 5 using four greenhouse gas emissions scenarios, we apply the statistical models to better understand their ability to predict water availability across ecological domains in both ungauged basins and under future climatological conditions.