Assessing and Validating the Ability of Machine Learning to Handle Unrefined Particle Air Pollution Mobile Monitoring Data Randomly, Spatially, and Spatiotemporally
Many epidemiological studies have evaluated the accuracy of machine learning models in predicting levels of particulate number (PN) and black carbon (BC) pollutant concentrations. However, few studies have investigated the ability of machine learning to predict the pollutant concentration with using unrefined mobile measurement data and explore the reliability of the prediction models. Additionally, researchers are moving away from using fixed-site data in favor of using mobile monitoring data in a variety of locations to develop hourly empirical models of particulate air pollution. This study compared the differences between long-term (daily average) and short-term (hourly average and 1 s unrefined data) model performance in three different classes of cross validation: randomly, spatially, and spatially temporally. This study used secondary data describing BC and PN pollutant levels in the rural location of Blacksburg (VA). Our results show that the model based on unrefined data was able to detect the pollutant hot spot areas with similar accuracy compared to the aggregated model. Moreover, the performance was found to improve when temporal data added to the model: the 10-fold MAE for the BC and PN were 0.44 μg/m3 and 3391 pt/cm3, respectively, for the unrefined data (one second data) model. The findings detailed here will add to the literature on the correlation between data (pre)processing and the efficacy of machine learning models in predicting pollution levels while also enhancing our understanding of more reliable validation strategies.