Concurrent spatiotemporal daily land use regression modeling and missing data imputation of fine particulate matter using distributed space-time Expectation Maximization

In this study, a spatiotemporal land use regression (LUR) model using Distributed Space-Time Expectation Maximization (D-STEM) software was developed. We trained the model using daily mean ambient particulate matter ≤ 2.5 μm (PM2.5) data measured hourly in 2015 at 30 regulatory monitoring network stations within the megacity of Tehran, Iran. Since a substantial amount of measured data were missing (48% of the total number of daily PM2.5 observations), we used the D-STEM to impute missing data and compared the missing imputation performance between different fitted models and the mean substitution method. We used h-block cross-validation (h-block CV) method in order to account for spatial autocorrelation in the model building and validation. In the imputation of missing data, the D-STEM LUR model had a mean absolute percentage error (MAPE) of 25.3%, outperforming the mean substitution method, which resulted in MAPE of 28.3%. The spatiotemporal R-squared was 0.73 and the average CV R-squared of 2-block and 5-block cross-validations was 0.60. These values were 0.68 and 0.47 when the spatial aspect of the LUR model was assessed, and 0.995 and 0.992 when the temporal aspect of the LUR model was assessed. This study demonstrated the competence of D-STEM software in spatiotemporal modeling, missing data imputation, and mapping of daily ambient PM2.5 at a very high spatial resolution (20 m x 20 m). These estimations are available for future research, especially for epidemiological studies on short- and/or long-term health effects of ambient PM2.5. Generally, we found D-STEM as a promising tool for spatiotemporal LUR modeling of ambient air pollution, especially for those models that rely on regulatory network monitoring stations with a considerable amount of missing data.