What controls the error structure in evapotranspiration models?

► We developed statistical models to describe the error structures of 3 different ET products. ► We use these models to conduct variable important and partial dependence analysis. ► Canopy indices EVI and LAI and flux magnitudes were strongest drivers of error. ► The decoupling coefficient and resistance terms were not strong drivers of error. ► Combined flux models or soil moisture estimation may improve ET model performance. Evapotranspiration models allow climate modelers to describe surface–atmosphere interactions, ecologists to understand the impact that global temperature change and increased radiation budgets will have on ecosystems, and farmers to decide how much irrigation to give their crops. Physically based algorithms for estimating evapotranspiration must manage a trade-off between physical realism and the difficulty of parameterizing key inputs, namely resistance factors associated with water vapor transport through the canopy and turbulent transport of water vapor from the canopy to ambient air. In this study we calculate predicted evapotranspiration at 42 AmeriFlux sites using two types of dedicated evapotranspiration models—one using physical resistances from the Penman–Monteith equation (Monteith, 1965) (Mu et al., 2007, 2011) and another based on the Priestley–Taylor (1972) equation, substituting functional constraints for resistances (Fisher et al., 2008). We analyze the structure of the residual series with respect to various meteorological and biophysical inputs, specifically Jarvis and McNaughton's (1986) decoupling coefficient, Ω, which is designed to represent the degree of control that plant stomata versus atmospheric demand and net radiation exercise over transpiration. We find that vegetation indices, magnitude of daytime fluxes, and bulk canopy resistance (rc)—which largely drives Ω—are strong predictors of patterns in model bias for all flux products. Though our analysis suggests a consistently negative relationship between Ω and mean predicted error for all evapotranspiration models, we found that vegetation indices and flux magnitudes were the most significant drivers of model error. Before addressing error associated with canopy resistance and Ω, refinements to existing models should focus on correcting biases with respect to flux magnitudes and canopy indices. We suggest a dual-model approach for backsolving rc (rather than estimating it from lookup tables and canopy indices), and increased attention to water availability,