Revealing the Covariation of Atmospheric O₂ and Pollutants in an Industrial Metropolis by Explainable Machine Learning

Abstract

In urban areas, atmospheric O 2 actively participates in the process of anthropogenic emissions and energy consumption. However, the covariation between atmospheric O 2 and the emitted pollutants has yet to be thoroughly explored. This study examines the covariations between atmospheric O 2 and pollutants in Lanzhou, a semi-arid industrial metropolis. A machine learning (ML)-based O 2 simulator coupled with a SHapley Additive exPlanation (SHAP) algorithm is established to explore and interpret their covariations under diverse conditions. Our findings indicate an increase of 16.3 ppm in the O 2 concentration associated with the atmospheric transport of natural dust particles during dusty weather events. This suggests that natural dust transport can mitigate the depletion of atmospheric O 2 caused by primary emissions and secondary formation of anthropogenic particulate matter. Furthermore, we identify a nonlinear relationship between the concentrations of O 2 and pollutant concentrations, which likely arises from their distinct diffusive abilities. The study highlights the unique pollution characteristics in a semi-arid urban downtown and demonstrates the ability of ML-based methodology to reproduce and interpret the environmental and anthropogenic impacts on the local carbon−oxygen balance.

Atmospheric O2 and CO2 measurements at a single height provide weak constraint on surface carbon exchange

“…To the best of our knowledge, no previous studies have reported such high deviations of ER atmos from ER forest , or ER atmos values higher than 2 for above forest canopy measurements as we found in Faassen et al (2023). Only Liu et al (2023b) found a difficult to explain nonlinear relationship between O 2 and other pollutants. While some differences between ER atmos and ER forest have been observed in previous studies, these differences typically fall within a range of 0.5 (Seibt et al, 2004;Ishidoya et al, 2015;Battle et al, 2019;Yan et al, 2023).…”