Abstract: Observation of air pollution levels at certain points in space and time is done by using mobile and static sensor networks. The values of air pollution levels at points where no measurements were made are mostly assumed by numerous types of interpolation between known values at measured points. The authors of this paper propose techniques for predicting air pollution levels in points in space where there are no measurements. The proposed techniques are based on the analysis of measurements from the sensor network that are affected by the same sources of pollution. Three approaches for identifying unknown air pollution sources by collecting measures from sensors mounted on public service vehicles are defined, implemented, and evaluated. The first approach can be treated as the optimization problem, the second approach is based on clustering in a multidimensional space and the third one is a fast and light method for a specific simplified case of the problem. The system is also implemented for a distributed computer cluster that applies machine learning algorithms over data streams for efficient estimation of dominant pollution sources in real-time.