An automated multi-model evapotranspiration mapping framework using remotely sensed and reanalysis data

Abstract

Estimating regional evapotranspiration (ET) is challenging in data-limited regions where a lack of in situ observations constrain model calibration and implementation. Here we developed an ensemble mean surface energy balance (EnSEB) modeling framework that is independent of any ground calibration and applied it in India to understand the magnitude and variability of ET in this agriculturally important region. EnSEB uses daily land surface temperature (LST) and vegetation biophysical inputs from Moderate Resolution Imaging Spectroradiometer (MODIS) and climatic information from the National Aeronautics and Space Administration (NASA) Modern-Era Retrospective Analysis for Research and Applications, version 2 (Merra-2) products and runs seven surface energy balance (SEB) algorithms to estimate ensemble mean ET or latent heat (LE) fluxes at a spatial resolution of 1 km × 1 km. Due to limited access to observed flux data, we conducted three different types of model evaluation: i) instantaneous LE validation using observed SEB fluxes from Bowen ratio energy balance (BREB) measurements in four agroecosystems, ii) annual and seasonal ET comparison with six global products at the regional scale, and iii) by closing the basin-scale monthly water budget (WB) for five large river basins (87,900–312,812 km2) in India. Validation with the BREB measurements revealed hourly EnSEB LE estimates to be within 2% of the observed LE (R2 = 0.57 and RMSE = 59 W m−2) and EnSEB was more accurate than any of the individual SEB models. Annual ET from EnSEB was positively correlated with six widely-used global ET products (r = 0.52–0.83, p-value < 0.001), but EnSEB captured the magnitude of ET in intensively irrigated regions much better. Basin-scale monthly WB miscloures were found to be between −1 and 9 mm month−1 from EnSEB ET estimates, which were better than those from the six global ET products. The gap filling method based on the constant ETrF (ET/reference ET) approach introduced some uncertainties in EnSEB, which presents room for future improvements. Overall, our results suggest that the automated and calibration-free multi-model EnSEB framework, which uses only remote sensing and readily available reanalysis data, has the ability to estimate ET with reliable accuracy in Indian agroecosystems. Such a framework could help us better understand and monitor the water cycle in regions where ground data are limited or non-existent.