Fusion of land use regression modeling output and wireless distributed sensor network measurements into a high spatiotemporally-resolved NO2 product.

Abstract

Land use regression modeling is a common method for assessing exposure to ambient pollutants, yet it suffers from very coarse temporal resolution. Wireless distributed sensor networks (WDSN) is a promising technology that can provide extremely high spatiotemporal pollutant patterns but is known to suffer from several limitations that put into question its data reliability. This study examines the advantages of fusing data from these two methods and obtaining high spatiotemporally-resolved product that can be used for exposure assessment. We demonstrate this approach by estimating nitrogen dioxide (NO2) concentrations at a sub-urban scale, with the study area limited by the deployment of the WDSN nodes. Specifically, hourly-resolved fused-data estimates were obtained by combining a stationary traffic-based land use regression (LUR) model with observations (15min sampling frequency) made by an array of low-cost sensor nodes, with the sensors' readings mapped over the whole study area. Data fusion was performed by merging the two independent information products using a fuzzy logic approach. The performance of the fused product was examined against reference hourly observations at four air quality monitoring (AQM) stations situated within the study area, with the AQM data not used for the development of any of the underlying information layers. The mean hourly RMSE between the fused data product and the AQM records was 9.3ppb, smaller than the RMSE of the two base products independently (LUR: 14.87ppb, WDSN: 10.45ppb). The normalized Moran's I of the fused product indicates that the data-fusion product reveals more realistic spatial patterns than those of the base products. The fused NO2 concentration product shows considerable spatial variability relative to that evident by interpolation of both the WDSN records and the AQM stations data, with significant non-random patterns in 74% of the study period.