Assessing and Correcting Bias in Gridded Reference Evapotranspiration over Agricultural Lands Across the Contiguous United States

Gridded reference evapotranspiration (ETo) data are widely used for agricultural water management and remote sensing ET (RSET) models, but biases can arise in agricultural regions where coarse-resolution meteorological inputs fail to capture local microclimates. We investigated biases in the gridMET ETo product across irrigated agricultural areas of the contiguous United States (CONUS) by comparing gridMET values with ETo calculated from 793 agricultural weather stations. We also used these stations to bias correct monthly gridMET ETo. Results show that gridMET systematically overestimates ETo by 10–20% at most cropland sites, while pockets of underestimation appear in some arid western regions, primarily due to wind speed bias. Overestimation of wind speed was the dominant driver of ETo bias, amplified by positive biases in solar radiation and maximum air temperature and negative biases in humidity (vapor pressure), whereas minimum temperature bias had a smaller effect. Regionally, the relative influence of these drivers varied: in some arid western climates, ETo bias was most closely linked to humidity and temperature errors, while in many other regions, including much of the eastern and coastal regions, wind speed and solar radiation biases were the dominant drivers. Comparison with 79 independent eddy covariance sites showed that the bias correction substantially improved monthly gridMET ETo accuracy at most of these stations. For example, the mean absolute error (MAE) in ETo was reduced at 80% of cropland sites across the CONUS. The bias-corrected ETo also improved accuracy in the three ETo-dependent OpenET RSET models (eeMETRIC, SIMS, and SSEBop), reducing MAE at 47–67% of cropland sites. Moreover, at the cropland eddy covariance stations, monthly mean bias error (MBE) of the three RSET models decreased by up to 9.5 mm/month (equivalent to 10% of the mean monthly eddy covariance ET), MAE was reduced by 1.7–3.5 mm/month, and root mean square error (RMSE) was reduced by 2.6–4.2 mm/month; though the OpenET ensemble accuracy metrics changed minimally overall. Across natural land cover types including forests, wetlands, grasslands, and shrublands, MAE and RMSE also improved for all RSET models. Our findings underscore the need to account for bias in gridded ETo data, which can arise from multiple factors in the underlying gridded meteorological data inputs and methods that we explore and discuss. We also discuss potential solutions, including alternative datasets and approaches for calculating ETo that could better capture climate conditions over irrigated land, and alternatives to the direct utilization of gridded ETo in water-limited regions for RSET modeling. Incorporating bias corrections for gridded ETo can reduce uncertainty in applications that directly use ETo and improve the reliability of RSET models and other water resource applications that depend on gridded ETo data.