Comprehensive evaluation of empirical algorithms for estimating land surface evapotranspiration

Many empirical algorithms for obtaining evapotranspiration (ET) from vegetation indices (VIs) have been developed, but there has been little work comparing these algorithms to each other or deriving coefficients for them using large data sets for training and validation. Twelve different vegetation index-based regression algorithms for retrieval of ET on a daily basis are reviewed and evaluated here. New coefficients have been derived for four of these algorithms using data from 181 Ameriflux and Fluxnet2015 sites and 1 km MODIS subsets centered at each site location. Algorithm validation with previously published and new coefficients was performed using one year of data from each Ameriflux and Fluxnet2015 site. There was a wide range of performance of these algorithms, with the median R2 by site in the 0.6 to 0.7 range, median root mean square error (RMSE) about 25 W/m2 and median bias within 10 W/m2. When algorithm coefficients were re-derived, the RMSE and bias of the worst-performing algorithms were largely reduced, but R2 was little changed. Agricultural and wetland sites had a low bias across most of the algorithms, and wetland sites had a higher RMSE. When several of the algorithms were re-tuned to obtain coefficients specific to each surface type, the biases of the agricultural and wetland sites were reduced to those more typical of other site types, and RMSE for agricultural and wetland sites was also reduced. The effects of linear interpolation of VIs to obtain daily LE and interpolation over periods of rapid VI change at agricultural sites were examined. No significant algorithm performance degradation was found in either case. It is recommended to use more detailed algorithms when possible, with inclusion of net radiation as a parameter along with VI at a minimum.