Abstract Fine particulate matter with aerodynamic diameters equal to or less than 2.5 micrometers (PM2.5) is a major component of air pollutants. The adverse effects of PM2.5 on public health have been well recognized. Numerous methods have been developed for ground-level PM2.5 concentration mapping by exploiting the advancements of machine learning and observational technologies. The currently available gridded long-term PM2.5 concentration datasets are typically at a spatial resolution of 1 km × 1 km (or 0.01° × 0.01°) and a temporal granularity of a year, which is insufficient for public health applications that require higher spatiotemporal resolutions. In this study, we propose a statistical approach to integrate random forests, a commonly used machine learning algorithm, and regression kriging, a recognized geostatistical method, for high-resolution mapping of ground-level PM2.5 concentrations. This approach jointly considers the heterogeneous geospatial variables that are closely related to the distribution of PM2.5, including meteorological factors, socioeconomic development activities, and topographic information. The integration of the machine learning and traditional geostatistical methods enables the effective modeling of the nonlinear relationships between the PM2.5 concentration and the predictor variables (via random forests) while accounting for the complex spatiotemporal effects of the variables (via kriging). Using this integrative approach, we produce a time-series (January 2014 to December 2014) monthly PM2.5 concentration dataset at a spatial resolution of 500 m for the contiguous United States. The advantages of the proposed approach are discussed and highlighted with a performance comparison with a commonly used land use regression method in ground-level PM2.5 concentration mapping.