Mapping Air Pollution with Google Street View Cars: Efficient Approaches with Mobile Monitoring and Land Use Regression

TLDR

Mobile monitoring campaigns can provide high‑resolution outdoor air pollution data. The study investigates how to reduce data needs for city‑wide air‑quality mapping using mobile monitors, comparing data‑only and predictive modeling strategies. The authors used two Google Street View cars equipped with 1‑Hz NO and BC sensors to collect data in Oakland, then applied Monte Carlo analyses to evaluate a data‑only approach and a land‑use regression‑kriging model for spatial mapping. The data‑only approach outperformed the LUR‑K model when 4–8 repeated drives per road segment were used, achieving cross‑validation R2 of 0.65 for NO and 0.43 for BC, whereas LUR‑K captured general trends but not full variability.

Abstract

Air pollution measurements collected through systematic mobile monitoring campaigns can provide outdoor concentration data at high spatial resolution. We explore approaches to minimize data requirements for mapping a city's air quality using mobile monitors with "data-only" versus predictive modeling approaches. We equipped two Google Street View cars with 1-Hz instruments to collect nitric oxide (NO) and black carbon (BC) measurements in Oakland, CA. We explore two strategies for efficiently mapping spatial air quality patterns through Monte Carlo analyses. First, we explore a "data-only" approach where we attempt to minimize the number of repeated visits needed to reliably estimate concentrations for all roads. Second, we combine our data with a land use regression-kriging (LUR-K) model to predict at unobserved locations; here, measurements from only a subset of roads or repeat visits are considered. Although LUR-K models did not capture the full variability of on-road concentrations, models trained with minimal data consistently captured important covariates and general spatial air pollution trends, with cross-validation R2 for log-transformed NO and BC of 0.65 and 0.43. Data-only mapping performed poorly with few (1–2) repeated drives but obtained better cross-validation R2 than the LUR-K approach within 4 to 8 repeated drive days per road segment.