Clean cycling in Amsterdam: development of a near-real time air pollution model from meteorology and traffic

Abstract

Recent research states that there is a need for cycling and walking route planning applications with air pollution as an input variable to a route planning algorithm. Several route planners exist, but none is specific about the methodology behind the route planner. This research attempted to develop a spatio-temporal model of air pollution that is suitable for use in an air-pollution based route planning application for the region of Amsterdam, using real-time air pollution as an input factor.

To this end, hourly open traffic intensity, meteorological and air pollution data from 2015 was used. Two statistical methods for modelling air pollution from traffic data and meteorological data were compared: multiple regression and the machine learning method random forest. Three air pollutants were considered (PM10, PM2.5 and NO2). The results highlight that the use of multiple regression on meteorological data is highly problematic, as many non-linear and multicollinear meteorological relations disrupt meaningful statistical analysis. The random forest technique led to a less interpretable outcome, but the R² for NO2, PM10 and PM2.5 were significantly higher, at +- 0.75 for all three pollutants. However, testing the random forest model on an air pollution dataset of November 2016 led to low explained variance for PM10 and PM2.5 (R² around 0.25), with NO2 scoring better (R² = 0.54). Variables with a high influence on measured hourly air pollution are traffic intensity and wind direction for all pollutants, with several other variables scoring high on different pollutant outcomes.

The application of machine learning algorithms in real-time mapping of air pollution needs further research. Further research is also needed on more fine-grained temporal data, as different time scales require different models. The development of a web application shows that the generated pollution values can be used well using the A* search algorithm. Urban citizens were found to be able to lower their exposure to air pollution by 27% on average pollution days, with a marginally longer route, highlighting the usability of low-pollution route planners.