Combined land-use and street view image model for estimating black carbon concentrations in urban areas

Abstract

In this study, we developed a novel land-use street view image random forest (LUSRF) model to estimate the equivalent black carbon (eBC) concentration based on land-use random forest (LURF) and street view imagery (SVI) models and compared their accuracy and precision in the urban city of Augsburg, Germany. The variables of the LUSRF model were constructed by combining LURF and SVI model variables (i.e., land-use, street scene, and meteorological factors). Stratified cross-validation (CV) was used to validate the model performance. Based on R2 and IA (Index of Agreement), LUSRF has superiority (average-R2: 0.73, average-IA: 0.91) compared to the LURF (average-R2: 0.52, average-IA: 0.81) and SVI model (average-R2: 0.68, average-IA: 0.89) in the urban city of Augsburg during the observed period. The main driving factors of the LUSRF model for BC estimation were different in heating and non-heating periods (i.e., elevation, the proportion of moving cars, and relative humidity for the non-heating period; and elevation, the proportion of building, and relative humidity for the heating period), which improves the estimation accuracy of eBC concentration and its sources. The model verification in other areas (i.e., suburban and small towns) further proved that the model has certain generalizability. Overall, the LUSRF model will provide insight for epidemiological studies in urban areas as a personal exposure assessment.