Investigating Population Exposure Assignment Methods for Air Pollution from Google Street View Polyline Data in Copenhagen, Denmark

Abstract

Listen

Translate article

BACKGROUND AND AIM: Google Street View (GSV) cars provide air pollution (AP) data across thousands of streets in multiple cities. Various methods exist for linkage of such vector data with populations. While rasterization or near-analysis are possible methods, multiple streets often surround residences; thus, a composite value can be assigned for geo-locations. We aimed to identify best geospatial method for exposure assignment from such data. METHODS: Long-term mean AP [ultra-fine particles (UFP), nitrogen dioxide (NO2), and black carbon (BC)] predictions across 30,312 streets (length = 15-60 m) were obtained from GSV-based mixed-effects LUR models developed for Copenhagen, Denmark. A near-analysis was used where Euclidean distances between each residence (out of ~77,000) and surrounding streets were calculated, nearest street was identified, and its AP values were assigned. Predictions were also assigned to mid-street centroid; using a systematic algorithm data were split to train (24,061; ~80%) and test sets (3,031; ~10%). Spatial averaging (SA), inverse distance weighting (IDW), ordinary kriging (OK), and natural neighbor (NN) models with multiple configurations for weighting and cell-size were developed. The coefficient of determination (R2) and RMSE were calculated on the test sets. RESULTS: The mean (SD) of UFP, NO2, and BC were, respectively, 14,120 (8,849) particles/cm³, 16.8 (8.3) μg/m³, and 1.1 (0.4) μg/m³. Overall, 9 SA, 27 IDW, 45 OK, and 3 NN models were developed. NN with a cell-size of 15m was the best performing model. The R2 and RMSE for NN on the test sets were, respectively, 0.92 and 2543 pt/cm3 for UFP, 0.87 and 3.1 µg/m3 for NO2, and 0.88 and 0.15 µg/m3 for BC. The Spearman correlation between residential predictions from NN and near-analysis assignment method was 0.97 for UFP, 0.95 for NO2, and 0.93 for BC. CONCLUSIONS: Although high correlation was observed for NN and near-analysis, the latter overestimated the concentrations.