Europe-wide air pollution modelling from 2000 to 2019 using geographically weighted regression

Abstract

BACKGROUND AND AIM Land-use regression (LUR) empirically relates air pollution concentrations to predictors such as traffic data, population, and land use data. Previously used regression methods in Europe wide models assumed spatially-fixed linear relationships between predictors and air pollution concentrations. Thus, to include spatial heterogeneity in relationships, we used geographically weighted regression (GWR) and the machine learning method Random Forest (RF), with the aim to further improve the accuracy of air pollution exposure estimates. METHODS We built separate LUR models for each year from 2000 to 2019 for NO2, O3, PM2.5 and PM10 using annual average concentrations from routine monitoring stations across Europe. The potential predictors included satellite retrievals, chemical transport model estimates and land-use variables. Supervised linear regression (SLR) was first used to select predictors, and then GWR estimated the potentially spatially-varying regression coefficients. We evaluated the model performance using five-fold cross-validation (CV) for each year and compared GWR with SLR and RF. RESULTS GWR explained measured concentrations generally well, with annual CV-R2 values of 0.62-0.67, 0.43-0.66, 0.48-0.71, 0.69-0.82, for NO2, O3, PM10, and PM2.5 respectively. GWR improved the R2 values compared to SLR by 4-7%, 7-15%, 7-14%, 0-6% and compared to RF by 1-4%, 0-10%, 0-11%, 7-19%, for NO2, O3, PM10, and PM2.5 respectively. Predictors selected and the spatially-varying coefficient values varied between years. CONCLUSIONS Including spatial heterogeneity using geographically weighted regression improved European air pollution LUR models. These models allow time-varying exposure-risk models. KEYWORDS Geographically weighted regression, spatial heterogeneity, land-use regression, air pollution, supervised linear regression, random forests