Accurate estimates of ambient air pollution levels are crucial for supporting health effect and health impact analyses. However, measurements from monitoring networks are spatially sparse, temporally incomplete, and often preferentially located. Over the past decade, many modeling approaches have been used to develop air quality products with complete spatial-temporal coverage, often by incorporating different sources of information (e.g., meteorology, emission sources, and satellite imagery). Performances of these approaches may vary across regions and time periods due to issues related data quality, data availability, and model assumptions. We describe a Bayesian ensemble approach that aims to combine predictions, as well as their uncertainties, from multiple models, where the ensemble weights are allowed to be covariate-dependent and spatially varying. The proposed approach is applied to estimate daily fine particulate matter air pollution at 1km spatial resolution in southeastern United States by considering predictions from Bayesian spatial-temporal models, Bayesian additive regression trees, and deterministic numerical model simulations.