Interpretable diurnal impacts on extreme urban PM2.5 concentrations of soil temperature, soil water content, humidity and temperature inversion

• Published in: Atmospheric research 2024-09, Vol.307, p.107500, Article 107500
• Main Authors: de Foy, Benjamin, Schauer, James J.
• Format: Article
• Language: English
• Subjects: air; Air pollution; air quality; data collection; eXplainable artificial intelligence; Generalized additive model; humidity; Interpretable machine learning; meteorology; Mixing height; model validation; PM2.5; pollution; soil temperature; soil water content; Temperature Inversion
• ISSN: 0169-8095; 1873-2895
• DOI: 10.1016/j.atmosres.2024.107500

Inhabitants of megacities around the world are suffering from severely unhealthy concentrations of PM2.5 as a result of intense emissions from numerous sectors, often combined with adverse meteorological conditions. Machine learning algorithms can be applied to multiyear datasets of hourly measurements in order to identify the main drivers causing intense pollution. In particular, Generalized Additive Models (GAM) provide interpretable associations of the interplay between emissions and meteorology. GAM simulations were developed for five dry seasons in Kolkata and Dhaka. In the model, soil temperature was associated with 21% of PM2.5 variability. Instead of 2-m humidity, model performance was improved by including air humidity at 1000 hPa and soil water content individually, with each accounting for around 6% of PM2.5 variability. Boundary layer heights have a significant impact on daytime concentrations, but the GAM output showed that temperature inversion intensity better characterized the stability of the nocturnal boundary layer and had a larger contribution to the PM2.5 variability. The GAM model could also identify interactions between parameters and showed that boundary layer height, temperature inversion intensity and air humidity had impacts that varied by time of day. By using GAM factors for winds at 100 m above the surface in combination with the Trajectory Cluster Contribution Function, the model estimated that local winds were associated with around 6% of variability whereas long range transport was associated with around 9% of variability. The analysis shows that there are no silver bullets for improving air quality and that adverse meteorology is making the problem harder to solve. Nevertheless, the results show that sustained efforts at controlling both local and regional sources will yield cleaner air that is greatly needed to improve the health of people living in large metropolitan areas. [Display omitted] •Generalized Additive Models yield interpretable associations of meteorology and hourly PM2.5•Temperature inversion intensity characterizes the stable nocturnal boundary layer•Boundary layer height is a weaker predictor of PM2.5 than temperature inversion intensity•To improve machine learning model include soil temperature and soil water content•Humid air and drier soils both lead to higher PM2.5 concentrations