Physically-Sound Air Pollution Regression Modelling with a Heterogeneous Wind Field

Abstract

Background/Aim: Detailed physical and chemical modelling of ambient pollutants transport is complex and requires extensive data and effort. Moreover, the output of such models often contain large errors. Alternatively, land use regression is often used but is limited as it lacks incorporation of the wind field – the single most influential parameter on pollutant dispersion. This work integrates a 1x1km data-driven meteorological model with a non-linear regression scheme and produces pollutant concentration maps at a high spatiotemporal resolution. We applied this model to estimate ambient NOx at the mountainous terrain of the Haifa Bay area, Israel. Methods: This work enhances the previously developed Optimized Dispersion Model (ODM), replacing the euclidean distance between grid points with the newly formulated Shortest Wind-Path Distance (SWPD). We used Dijkstra’s algorithm to determine the wind direction affected SWPD between each two grid points at each time-point. We use the calibrated model to estimate the relative influence of each sector (traffic, industry) on the observed NOx levels across the whole study area. Data on point source emissions was obtained from the Israeli Pollutant Release and Transfer Registry, while detailed traffic emissions proxies were obtained from aggregated vehicle-fleet GPS tracking. Results: Complete leave-one-out cross-validation showed that the new model performs better than a previous ODM version and a benchmark geo-interpolation model. Model performance improved greatly for increasing time-averaging windows (i.e. the mean spatial Pearson correlation was 0.3 based on half-hourly measurements and 0.8 for the yearly mean). Traffic was shown to have the major contribution to observed ambient NOx in the study area, although industrial NOx emissions are much greater than traffic-induced emissions. Conclusions: The new data-driven yet physically sound ODM air quality model enables much better exposure estimation for epidemiological research. Unlike other models, the new model can be applied in areas characterized by a heterogeneous wind field.