On spatial prediction of urban air pollution

Air pollution continues to draw global attention, and still causes adverse environmental issues and health effects. Questions about how air pollution evolves spatially are also still unsolved. Many air quality monitoring stations are deployed in several countries to give insight about air quality. However, it is quite frequent for these stations to go out of order during their long life time. These incidents may lead to a significant loss of pollution data. To mitigate this issue, we propose to leverage spatial correlation between pollution monitoring stations to predict the lost data. To reduce the complexity of the prediction model, for each monitoring station, we identify the best set of other stations to use in the prediction model. Correlation-based station selection is shown to outperform distance-based station selection and provides an R-2 above 0.8 when applied to the Airparif datasets.