Development and validation of land use regression models for ultrafine particles in Augsburg and Regensburg, Germany

Ultrafine particles (UFP) are suspected to have a high toxic potential, but evidence from long-term epidemiological studies remains sparse since highly spatially resolved UFP data is lacking. We modelled long-term annual average total particle number concentration (PNC) as indicator for UFP for two middle-sized German cities (Augsburg and Regensburg) and their surroundings, which are part of the German National Cohort (NAKO), for subsequent linkage with health data. Supervised land use regression (LUR) models were developed for Augsburg, combining two previous measurement campaigns (monitoring sites: 2014/15: N = 20 and 2017: N = 6) and spatial predictors. To account for the time difference and repeated monitoring sites, we applied a generalized additive model (GAM) and a mixed model (MM). Models were internally validated using leave-one-out cross-validation (LOOCV). We transferred the models to the Regensburg region and externally validated our predictions using in-situ measurements carried out in 2020/21 at six monitoring sites. For both approaches, models showed highly adjusted explained variance and LOOCV R 2 (GAM: 0.90 and 0.76; MM: 0.91 and 0.86). Similar predictors were selected, mainly indicators for road network and industrial areas. The external validation showed good agreement of measured and predicted PNC with Spearman correlation coefficient r = 0.75 (GAM) and 0.86 (MM), though both models tended to underestimate the concentrations. The two LUR models resulted in similar predictions and captured intra-city spatial patterns and city-rural gradients well. The Augsburg models could be effectively transferred to Regensburg since the study regions featured similar characteristics. To evaluate the predictive capability in novel study areas, external validation measurements are recommended. • High-resolution spatial ultrafine particle data lacking for environmental health analyses. • Land use regression models developed for Augsburg and transferred to Regensburg. • Models performed very well and showed similar spatial patterns and contrasts. • External validation showed good agreement, but underestimation at one traffic site. • Validation important to support epidemiological data analyses.