Remotely-sensed parameters, such as aerosol optical depth (AOD), have been used extensively in estimating ambient air pollution. Satellite-based products have the advantages of fine spatial resolution and wide spatial coverage, but they are subject to non-random missingness due to cloud over or high surface reflectance that are associated with pollution levels. The amount of missingness is compounded by the lack of daily air quality measurements, especially for PM2.5 constituents. Hence, for long-term air pollution studies, satellite-based estimates need to be gap-filled to avoid biased exposure and health effect estimates. We describe a Bayesian ensemble approach that utilizes monitoring measurements, satellite-based AOD, and numerical model simulations. Our approach will incorporate prediction uncertainties associated with each ensemble member, and allow for spatial-temporal covariates and spatial dependence in ensemble weights. In a case study for Southeastern US, we found that the proposed approach not only provides complete coverage for estimating daily PM2.5 at 1km spatial resolution but it also improves prediction performance.