The wide availability of data observed over time and space, due to widespread collection of network and inexpensive geographical information systems, has stimulated many studies in a variety of disciplines. These huge collections of data often contain possibly interesting and valuable information, which has raised the demand in spatio-temporal data analytic approaches. This talk will highlight some intelligent semiparametric regression models that are sufficiently flexible to incorporate the nonstationary and heterogeneous features of spatio-temporal data. An advanced spatial smoothing technique will be introduced to solve the problem of "leakage" across the complex domains where many conventional spatial analysis tools suffer from. To demonstrate the efficiency of our method, we perform a spatio-temporal analysis of fine particulate matter (PM2.5) in air pollution, which has been identified as one of the major factors strongly associated with increased cardiovascular disease and other various health problems. The results demonstrate that there are substantial benefits in modeling the spatio-temporal nonstationary and heterogeneous meteorological effect on PM2.5 concentrations.