Modeling the spatio-temporal distribution and characteristics of particulate matter remains a focal point of research. In the present study we model the spatio-temporal structure of with and without a spatial trend made of environmental covariates. (i) Spatio-temporal interpolation without accounting for covariates is done by using ordinary space-time kriging. (i) To include covariates into the model we propose the following methodology, a combination of methods that has not been investigated before: generalized additive regression is employed to capture the effects of the covariates by partitioning the output into a trend and a residual component. Furthermore, the unknown trend components at ungauged locations are estimated by spatial artificial neural networks and the corresponding residual components are predicted by means of ordinary space-time kriging based on a nested spatio-temporal covariance function that is optimized by a particle swarm algorithm. The results of both methods (i and ii) are compared by means of cross validation and it is found that the new methodology which takes the covariates into account performs significantly better, in particular, it yields a smaller mean squared error.