Hourly LUR modeling of hyperlocal NO2 using mobile monitoring data

BACKGROUND:

Distinguished from the real-time forecast task, the hourly mapping in this work stands for spatiotemporal interpolations of long-term (e.g., annual) average concentrations of air quality in hourly intervals. These long-term-averaged hourly maps facilitate environmental epidemiology studies by enabling dynamic environmental exposures assessment based on human time activities. Over the past decade, several mobile campaigns have been conducted where high-frequency sensors were mounted on vehicles roaming around the entire studied region. Since not all locations can be measured repeatedly for every hour, there are rarely previous studies using opportunistic mobile measurements to reconstruct long-term NO₂ hourly maps. In this work, we evaluated the merit of land-use regression (LUR) models that use mobile measurements paired with land-use and traffic predictors to interpolate hourly air pollution concentrations at fine spatial resolutions.

METHOD:

We monitored 1-second NO₂ concentrations in Amsterdam with two Google StreetView cars, from 8:00 to 20:00 on weekdays from May 2019 to March 2020 (5.7 Million measurements). These measured GPS points were aggregated into 50m road segments and divided into one-hour intervals. Using this hourly mobile data as the response, we explored two spatiotemporal LUR models, namely ST-Kriging (the spatiotemporal version of kriging methods) and GTWR (Geographical and Temporal Weighted Regression), and two spatial LUR models implemented separately in each hour, namely RF_LUR (a LUR model based on the random forest) and LSR (Stepwise Linear Regression). Model performance was assessed by averaging measurements over the same period collected from independent routine monitoring stations in Amsterdam (RIVM, n=9).

RESULT:

The hourly averaged mobile measurements of NO₂ across the city varied from 40 (rush hours) to 28 ug/m³ (non-rush hours). Routine measurements (RIVM) showed significantly different patterns in road types (major vs residential roads) and seasons (winter vs summer). Therefore, the spatiotemporal models were trained separately for these four scenarios and then merged their predictions into the final maps. GTWR captured more accurate spatiotemporal correlations than Kriging methods under the limitation of opportunistic mobile data and temporally static covariates (ST-Kriging: R² = 0.35, MAE, RMSE = 9.46, 12.14 ug/m³, GTWR: R² = 0.50, MAE, RMSE = 6.07, 7.64 ug/m³). Better overall accuracy and more smoothing distributions in both space and time were captured by the spatiotemporal models as compared to spatial models separated in each hour (LSR: R² = 0.47, MAE, RMSE = 6.48, 8.04 ug/m³; RF_LUR: R² = 0.33, MAE, RMSE = 10.33, 14.74 ug/m³). The spatiotemporal distribution of NO₂ predictions was found to strongly follow the intra-urban commuting pattern.

 CONCLUSION:

The spatiotemporal LUR model is able to capture spatiotemporal correlations hidden in opportunistic mobile measurements. The reconstructed spatiotemporal maps can be broadly applied to estimate human exposure to NO₂ considering time-activity patterns.