geeSEBAL-MODIS: Continental-scale evapotranspiration based on the surface energy balance for South America

•A novel methodology was developed for the application of the geeSEBAL model to very large scales, overcoming the issues associated with domain size selection.•Development of a 20-year (2002–2021) actual evapotranspiration dataset with complete spatial coverage for South America, providing 8-day evapotranspiration data at 500 m, hereon named geeSEBAL-MODIS.•The dataset was compared to eddy covariance measurements and water balance data across the continent.•geeSEBAL-MODIS yielded higher accuracy and performance than existing global ET datasets at both field and basin scales. Monitoring actual evapotranspiration (ET) is critical for the accurate assessment of water availability and water resources management, especially in areas with dry climates and frequent droughts. The Surface Energy Balance Algorithm for Land (SEBAL) has been used over several land and climate conditions, and is able to estimate ET at field scale with high accuracy. However, model complexity and subjective parameterization have hindered its operationalization, until the recent development of the geeSEBAL model, which implements the SEBAL model on the Google Earth Engine platform. Here, we present a unique methodology for a continental-scale application of SEBAL, called geeSEBAL-MODIS, that employs novel land surface temperature normalization techniques, enabling the application of contextual ET models to very large scales. We introduce a dynamic ET dataset for the entire South American continent, between 2002 and 2021, at 500 m spatial and 8 days temporal resolution. The satellite-based data were compared against daily ET measured at 27 flux towers as well as water balance-based annual ET from 29 large river basins. geeSEBAL-MODIS data were also compared to eight state-of-the-art global ET datasets. At local scale, geeSEBAL-MODIS demonstrated a satisfactory performance (correlation (r) = 0.65, Kling-Gupta Efficiency (KGE) = 0.64, Mean Absolute Error (MAE) = 0.83 mm day−1 (24.7 %) and Root Mean Squared Error (RMSE) = 1.07 mm day−1 (31.8 %)), with negligible bias. At basin scale, geeSEBAL-MODIS generally underestimated ET (bias = -85 mm year−1, r = 0.65, KGE = 0.47, MAE = 107 mm year−1 (10.1 %) and RMSE = 137 mm year−1 (12.9 %)). Compared to other global datasets, geeSEBAL-MODIS demonstrated better performance over multiple South American biomes, climates and land cover types. The developed dataset also provides lower errors (local monthly RMSE = 23.0 mm month−1 and basin annual RMSE = 1