Continental-scale Mapping of 30-meter Evapotranspiration Using Machine Learning

Energy balance modeling from moderate-resolution satellite imagery is quickly becoming the de facto standard for evapotranspiration (ET) measurements over large spatial extents, offering spatially continuous estimates as an alternative to sparse and discontinuous in-situ observations. Remote sensing (RS) models offer novel and robust datasets for the scientific community, but are limited temporally by satellite data availability and spatially by gridded input parameter requirements. This study trained random forest and linear regression models to predict annual 30-m ET from the Operational Simplified Surface Energy Balance (SSEBop) model at over 2 million km2 of the conterminous United States (CONUS) land surface from 2010-2017. To produce models capable of ET prediction outside the timeframe of satellite data availability, input features were restricted to common climate, landcover, and geophysical datasets that are independent from RS observations. The random forest model universally performed better than the linear model, with average normalized root mean squared error (RMSE) of 19% over the CONUS. In the eastern US, random forest returned RMSE of 13% over all land cover classes and 10% over forested land, compared to errors in western sites of 28% over all land cover classes and 15% over forests. Larger errors in western sites were disproportionately impacted by extensive shrub vegetation and barren surfaces where the model struggled to accurately capture SSEBop ET. Input features of longitude, land cover, climate, latitude, elevation, topographic diversity, and surface soil bulk density were shown to influence the random forest model most heavily. Better performance of random forest over linear regression models is attributed to their ability to support non-linear relationships between ET and longitude, precipitation, and elevation. The random forest model developed here allows for future use cases such as hindcasting and forecasting 30-m ET, evaluating the effects of land cover change on ET and basin hydrology at a variety of scales, and rapid field-scale ET approximation outside the CONUS.