Aerodynamic roughness length (z0) fundamentally affects land surface momentum loss and wind resource simulation, but ground truth data of z0 are sparse in space, causing z0 datasets used in atmospheric models are empirically estimated from land cover types through a look-up table. In this study, we derived z0 values from 101 anemometer towers in China. Taking them as ground truth, we show that existing gridded z0 datasets determined from either a look-up table or a machine-learning method contain considerable uncertainty and fail to capture the variability of z0 within each land cover type, although the latter performs better. Even for the widely-used ERA5, its z0 is overestimated in wind-rich regions of China, causing an underestimation of near-surface wind speed. This highlights the urgency to improve z0 data in atmospheric models. Current rapidly expanding anemometer towers may substantially enrich z0 truth data and thus provide potential to improve wind resource modeling.