A supervised machine learning approach to classify traffic-derived PM sources based on their magnetic properties

Environmental magnetism techniques are increasingly used to map the deposition of particulate pollutants on any type of accumulative surfaces. The present study is part of a collective effort that begun in recent years to evaluate the efficiency of these techniques involving a large range of measurements to trace the source signals. Here we explore the possibilities provided by the very simple but robust k-near-neighbors algorithm (kNN) for classification in a source-to-sink approach. For this purpose, in a first phase, the magnetic properties of the traffic-related sources of particulate matter (tire, brake pads, exhaust pipes, etc.) are used to parameterize and train the model. Then, the magnetic parameters measured on accumulating surfaces exposed to a polluted air as urban plant leaves and passive filters are confronted to the model. The results are very encouraging. The algorithm predicts the dominant traffic-related sources for different kinds of accumulative surfaces. The model predictions are generally consistent according to the sampling locations. Its resolution seems adequate since different dominant sources could be identified within one street. We demonstrate the possibility to trace traffic-derived pollutants from sources to sinks based only on magnetic properties, and to eventually quantify their contributions in the total magnetic signal measured. Because magnetic mapping has a high-resolution efficiency, these results open the opportunity to complement conventional methods used to measure air quality and to improve the numerical models of pollutant dispersion.