Comment on Golkar and de Oliveira (2026): Missing eddy covariance observations undermine reported OpenET biases

Source abstract

In a recent study published in Agricultural and Forest Meteorology, Golkar and de Oliveira (2026) reported systematic overestimation of evapotranspiration (ET) by OpenET across US climate zones and substantial improvements through machine learning bias correction. Here, we identify substantial inconsistencies in their eddy covariance (EC) reference data that challenge these conclusions. We first identified unusually low daily mean observed ET relative to previously reported values at the same AmeriFlux sites that is traceable to numerous zero-valued observations in their comparison scatterplot, and extended periods of exactly zero observed ET in their time series. These inconsistencies led us to hypothesize that missing latent heat flux observations were treated as zero, systematically lowering the reference EC ET measurements. Our independent reconstruction of daily ET from the original AmeriFlux datasets reproduced the reported site-mean ET within 0.04 mm d⁻¹ at 20 of 22 sites when missing observations were treated as zero, providing compelling quantitative evidence that missing data were effectively incorporated as zero ET in the original analysis. We further find that the reported ET statistics are reproducible only when omitting the energy balance correction described in the original study. These inconsistencies substantially inflate apparent OpenET overestimation and call into question the reported effectiveness of the proposed bias correction. Consequently, the study's conclusions regarding seasonal and climate-dependent correction mechanisms likely largely reflect artifacts in the reference observations rather than reflecting deficiencies in OpenET. Our findings demonstrate that errors in processing observational reference data can fundamentally alter the evaluation of satellite ET performance, underscoring the importance of rigorous quality control and reproducible analysis to distinguish genuine model biases from observational artifacts and reliably assess the effectiveness of proposed ML corrections.