Bayesian Estimation of Land-Use Regression Model Parameters for Air Pollution Mapping by Assimilating Satellite and Ground-Based Observations

Abstract

Land-use regression (LUR) has been used to investigate spatial variations in ambient concentrations of nitrogen dioxide (NO2). Satellite data has been included in LUR as a predictor variable and shown improvement in predicting concentrations. However, satellite data reflecting spatial distributions of NO2 should be used to estimate LUR parameters and be assimilated in LUR. Additionally, limited studies have investigated the prediction uncertainty in LUR. Therefore, this study not only optimizes LUR models through assimilating both satellite and ground-based observations but also evaluates uncertainty in model structure, parameters and measurements implicitly through the Generalized Likelihood Uncertainty Estimation (GLUE) method and explicitly through the particle filter (PF). The satellite data was collected from the TROPOspheric Monitoring Instrument (TROPOMI). Three calibration settings, using both satellite and sensor data, only satellite data and only sensor observations, were used to examine the contribution of the observations and to evaluate the resulting uncertainty. From the GLUE result, the median predictions of the optimal models calibrated by the two data sets explained 59.7% of annual average ambient NO2 variation, and the sensor-calibrated model performed similarly (R2 =0.614). It indicates that assimilating satellite data with a 12.5-km resolution does not improve optimizing the parameters that provide specific pollution emission sources at a high resolution (25-meter) but can capture the regional concentrations because of the large difference in their spatial resolutions. The fitting between predicted surface concentrations and the TROPOMI tropospheric VCDs gives a new perspective in relating tropospheric VCDs and ambient concentrations, giving an alternative method for deriving the column-to-surface ratio in future study.