Air pollution is a major global environmental problem that affects health. In view of this, it is important to improve the prediction method of air pollutant concentrations to obtain accurate exposure estimates. Currently, land use regression (LUR) models are widely used to predict the fine-scale spatial variation of air pollutants. However, most of previous studies used linear regression methods such as generalized linear models (GLM) with less applicability to fit LUR models. Considering the potential nonlinear relationship between predictor variables, this study adopted generalized additive models (GAM) to derive LUR models of air pollutants (including PM_2.5, PM₁₀, CO, NO₂, SO₂, and O₃) and air quality index (AQI) in Beijing with annual resolution. These models were based on routine monitoring data from 35 national regulatory monitoring sites and combined with a set of predictor variables such as land-use type, traffic and industrial emissions, population density, and meteorological factors. Results indicated that compared with traditional methods, the GAM approach significantly improved the performance of LUR models with explanatory power adjusted R² levels ranging from 70 to 90%, and the cross-validation analysis also showed high prediction accuracy of the GAM approach. Besides, this approach emphasized the importance of meteorology in predicting air pollutant concentrations and AQI values. Generally, this study provides a feasible way to determine exposure assessment in heavily polluted cities and future support for long-term environmental epidemiological studies.