Background/Aim: Exposure to nitrogen dioxide (NO2) pollution has been associated with a range of adverse health outcomes. Annual average pollutant concentrations are often used to estimate exposure, however, these estimates can be imprecise due to difficulty modelling spatial patterns at the resolution of neighbourhoods (e.g. a scale of tens of metres) rather than at a coarse scale (around several kilometres resolution). The objective of this study was to derive improved estimates of neighbourhood level pollutant concentrations for health studies by blending air pollutant measurements with modelled predictions using Bayesian analyses. Methods: We implemented the Bayesian Maximum Entropy (BME) approach to blend 'hard' data (considered precise) and probabilistic 'soft' data (with uncertainty defined using informative priors). We compiled and harmonised NO2 data from fixed-site monitors, chemical transport models, and satellite-based land use regression models to estimate neighbourhood level annual average NO2 concentrations in Sydney, Australia. The spatial model integrated the different underlying probabilities to produce a posterior probability density function of neighbourhood exposures. The mean of the posterior density was our estimate of NO2 exposure. Results: Estimated annual average concentrations from the BME model ranged from 3 to 35 ppb.Validation using independent data from a separate set of samples (using passive sampling methods) showed improvement, with Root Mean Squared Error (RMSE) of 2.6 ppb compared with the land use regression (2.8 ppb) and chemical transport model (3.1 ppb). Conclusions: Our study implemented state-of-the-art methods for exposure assessment and demonstrated an improvement in validation test statistics. In future work we will explore the impact of these improvements on exposure misclassification bias when used in epidemiological analyses of the impact of air pollution on health.