Gap-filling usually accounts for a large source of uncertainties in the annual CO₂ fluxes, though gap-filling CO₂ fluxes is challenging at dryland sites due to small fluxes. Using data collected from a semiarid site, four machine learning methods are evaluated with different lengths of artificial gaps. The artificial neural network and random forest methods outperform the other methods. With these methods, uncertainties in the annual CO₂ flux at this site are estimated to be within 16 g C m⁻².