Leveraging Google Place of Interest (POI) Data, Crowdsourcing, and Machine Learning to Predict Urban NO2 Concentrations for the Contiguous U.S.

Abstract

Background: Land Use Regression (LUR) has been extensively applied to model intra-urban variations of air quality across the US. Predictor variables are typically derived from government sources (e.g., US Census) and do not include detailed information on land use or urban form. We developed LUR models that combine enhanced data on urban form and land use with machine learning to compare to traditional LUR models.Methods: Our LUR models are based on data from three sources: (1) a crowdsourced measure of urban form, (2) land use information derived from Google Place of Interest (POI) data, and (3) satellite observations of air quality. We developed models using US EPA NO2 monitoring data (n=426 locations) during 2010-2015. The measure of urban form (Local Climate Zones [LCZs]) was obtained using a crowdsourcing platform (Amazon Mechanical Turk) to allow multiple users (n=10) to classify satellite imagery into 17 categories of urban development (e.g., compact; sprawl). We web-scraped Google POI data (e.g., gas stations; restaurants) at various buffers (100-1,000 meters) to add detailed information on land use. We also included satellite observations of NO2 abundance. Using only these variables, we developed LUR models using both conventional (stepwise regression) and machine learning (e.g., bagging; gradient boosting).Results: Model fit was better for the machine learning models (gradient boosting; R2: 0.59; mean absolute error [MAE]: 2.46) as compared to the conventional models (i.e., stepwise regression; R2: 0.39 and MAE: 3.35) when using only the LCZ urban form measure and satellite NO2 abundance. Adding Google POI data improved model fit. For gradient boosting model-R2 increased to 0.65 (MAE: 2.28); model-R2 increased to 0.61 for stepwise regression (MAE: 2.34).Conclusions: Our work suggests that using generalizable data on urban form (e.g., Google POI; LCZs) to develop LUR models may produce similar results to models developed using conventional data sources.