Abstract
In recent years, there has been growing interest in developing air pollution prediction models to reduce exposure measurement error in epidemiologic studies. However, efforts for localized, fine-scale prediction models have been predominantly focused in the United States and Europe. Furthermore, the availability of new satellite instruments such as the TROPOsopheric Monitoring Instrument (TROPOMI) provides novel opportunities for modeling efforts. We estimated daily ground-level nitrogen dioxide (NO(2)) concentrations in the Mexico City Metropolitan Area at 1-km(2) grids from 2005 to 2019 using a four-stage approach. In stage 1 (imputation stage), we imputed missing satellite NO(2) column measurements from the Ozone Monitoring Instrument (OMI) and TROPOMI using the random forest (RF) approach. In stage 2 (calibration stage), we calibrated the association of column NO(2) to ground-level NO(2) using ground monitors and meteorological features using RF and extreme gradient boosting (XGBoost) models. In stage 3 (prediction stage), we predicted the stage 2 model over each 1-km(2) grid in our study area, then ensembled the results using a generalized additive model (GAM). In stage 4 (residual stage), we used XGBoost to model the local component at the 200-m(2) scale. The cross-validated R(2) of the RF and XGBoost models in stage 2 were 0.75 and 0.86 respectively, and 0.87 for the ensembled GAM. Cross-validated rootmean-squared error (RMSE) of the GAM was 3.95 μg/m(3). Using novel approaches and newly available remote sensing data, our multi-stage model presented high cross-validated fits and reconstructs fine-scale NO(2) estimates for further epidemiologic studies in Mexico City.