Evaluation of land use regression models for NO₂ and particulate matter in 20 European study areas: The ESCAPE project

Land use regression models (LUR) frequently use leave-one-out-cross- validation (LOOCV) to assess model fit, but recent studies suggested that this may overestimate predictive ability in independent data sets. Our aim was to evaluate LUR models for nitrogen dioxide (NO₂₎ and particulate matter (PM) components exploiting the high correlation between concentrations of PM metrics and NO₂. LUR models have been developed for NO₂, PM_2.5 absorbance, and copper (Cu) in PM₁₀ based on 20 sites in each of the 20 study areas of the ESCAPE project. Models were evaluated with LOOCV and "hold-out evaluation (HEV)" using the correlation of predicted NO₂ or PM concentrations with measured NO₂ concentrations at the 20 additional NO₂ sites in each area. For NO₂, PM_2.5 absorbance and PM₁₀ Cu, the median LOOCV R²s were 0.83, 0.81, and 0.76 whereas the median HEV R ² were 0.52, 0.44, and 0.40. There was a positive association between the LOOCV R² and HEV R² for PM_2.5 absorbance and PM₁₀ Cu. Our results confirm that the predictive ability of LUR models based on relatively small training sets is overestimated by the LOOCV R²s. Nevertheless, in most areas LUR models still explained a substantial fraction of the variation of concentrations measured at independent sites.