Accurate measurements of ground-level PM2.5 (particulate matter with aerodynamic diameters equal to or less than 2.5 μm) concentrations are critically important to human and environmental health studies. In this regard, satellite-derived gridded PM2.5 datasets, particularly those datasets derived from chemical transport models (CTM), have demonstrated unique attractiveness in terms of their geographic and temporal coverage. The CTM-based approaches, however, often yield results with a coarse spatial resolution (typically at 0.1° of spatial resolution) and tend to ignore or simplify the impact of geographic and socioeconomic factors on PM2.5 concentrations. In this study, with a focus on the long-term PM2.5 distribution in the contiguous United States, we adopt a random forests-based geostatistical (regression kriging) approach to improve one of the most commonly used satellite-derived, gridded PM2.5 datasets with a refined spatial resolution (0.01°) and enhanced accuracy. By combining the random forests machine learning method and the kriging family of methods, the geostatistical approach effectively integrates ground-based PM2.5 measurements and related geographic variables while accounting for the non-linear interactions and the complex spatial dependence. The accuracy and advantages of the proposed approach are demonstrated by comparing the results with existing PM2.5 datasets. This manuscript also highlights the effectiveness of the geographical variables in long-term PM2.5 mapping, including brightness of nighttime lights, normalized difference vegetation index and elevation, and discusses the contribution of each of these variables to the spatial distribution of PM2.5 concentrations.