Land Use Regression models for Ultrafine Particles

Abstract

TPS 651: Air pollution exposure modeling 1, Exhibition Hall, Ground floor, August 27, 2019, 3:00 PM - 4:30 PM Background: A lack of routine monitoring for ultrafine particle (UFP) means that bespoke monitoring (e.g. mobile measurements on vehicles or repeated short-term monitoring over a spatially distributed network of sites) has to be undertaken to develop land use regression (LUR) models. This is especially challenging in mega-cities due to the large spatial area. Methods: As an alternative to monitoring across a mega-city, we developed single- and two-district models in London, and then tested the transferability (i.e. generalisability) of models between districts. In each district, repeated 30-min monitoring was undertaken at 40 sites in three different seasons (2016-2018) to estimate annual mean UFP (particles cm-3). We constructed ten single- and two-district LUR models by allocating monitoring sites to one of ten groups and then used nine groups of sites to develop each model, iteratively, with the remaining group held-out for model evaluation (e.g. groups 1-9 to predict for group 10; groups 2-10 to predict for group 1, etc.). We assessed model performance by comparing R2 and the variables selected, and pooled the 10% of sites held-out each time to produce an overall R2 to assess model robustness. The transferability of models was tested by applying single- and two-district models to the other districts. Results: For the ten models for single districts, R2 ranged from 60% to 73% and model variables were similar and all included nearby road traffic. Model performance and structure was different between areas. R2 for two districts ranged from 50% to 68%. Pooled holdout validation had R2 values of 37%- 59%. Performance in transferring models to other districts within London was overall lower with R2 ranging from 10% to 39%. Conclusion: UFP LUR models may perform well within the confines of a monitoring network but transferring models within a city may have substantially lower performance.