A hybrid kriging/land-use regression model to assess PM2.5 spatial-temporal variability.

Proximate pollutant data can provide information for land-use predictors in LUR models, when coupled with spatial interpolation of ambient pollutant measurements, may provide better pollutant predictions. This study applies a hybrid kriging/LUR model to assess the spatial-temporal variability of PM2.5 for Taiwan. Using PM2.5 concentrations at 71 EPA monitoring stations from 2006 to 2011, pollutant gradient surfaces were spatially interpolated using a leave-one-out ordinary kriging method based on "n-1" observations. The predicted concentration level of the targeted site was then extracted from the generated kriging map and adopted as a variable in LUR modelling. Annual and monthly resolutions of LUR models were developed to assess the effects by incorporating kriging-based estimates into pollutant predictions. The R2 obtained from conventional LUR procedures was 0.66 and 0.70 for annual and monthly models, respectively, whereas models using the hybrid approach showed better explanatory power (R2 of annual model: 0.85; R2 of monthly model: 0.88). Moreover, kriging-based PM2.5 estimates were the most important factor in the resultant models according to the dominant partial R2 of 0.82 and 0.7 in monthly and yearly models. Cross-validation and external data verification showed similar results, demonstrating robustness of the proposed approach. Using governmental pollutant observations is usually publicly available for most areas, this method provides an efficient mean to better assess PM2.5 spatial-temporal variations and predicts levels for nonmonitored areas.