Ammonia (NH3) released to the atmosphere leads to a cascade of impacts on the environment, yet estimation of NH3 volatilization from cropland soils (VNH3) in a broad spatial scale is still quite uncertain in China. This mainly stems from nonlinear relationships between VNH3 and relevant factors. On the basis of 495 site-years of measurements at 78 sites across Chinese croplands, we developed a nonlinear Bayesian tree regression model to determine how environmental factors modulate the local derivative of VNH3 to nitrogen application rates (Nrate) (VR, %). The VNH3-Nrate relationship was nonlinear. The VR of upland soils and paddy soils depended primarily on local water input and Nrate, respectively. Our model demonstrated good reproductions of VNH3 compared to previous models, i.e., more than 91% of the observed VR variance at sites in China and 79% of those at validation sites outside China. The observed spatial pattern of VNH3 in China agreed well with satellite-based estimates of NH3 column concentrations. The average VRs in China derived from our model were 14.8 ± 2.9% and 11.8 ± 2.0% for upland soils and paddy soils, respectively. The estimated annual NH3 emission in China (3.96 ± 0.76 TgNH3·yr(-1)) was 40% greater than that based on the IPCC Tier 1 guideline.