aceBeyond concentration-based analysis: explainable AI identifies environmental settings influencing urban PM 10 variability
Abstract: This study applies an explainable artificial intelligence framework to investigate PM10 variability using routine regulatory air-quality data from a single monitoring station, targeting data-limited conditions. A four-year dataset (2020-2023) of PM10, PM2.5, NO2, SO2, O3, and meteorological predictors was analyzed using ensemble machine-learning models with metaheuristic hyperparameter optimization. The best-performing model achieved high predictive performance (R2 = 0.913), supporting model-based interpretation…
Assessment of air quality using a cloud model method by Xu, Qingwei; Xu, Kaili. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0
Researchers applied an explainable AI framework to four years of routine air-quality and meteorological data from a single urban monitoring station to move beyond concentration-only analysis of PM10. The best ensemble model reached R2=0.913, and SHAP-based clustering revealed ten recurrent environmental settings linked to enhancement, reduction, or transitional PM10 behavior.
The approach matters because it extracts interpretable drivers from standard regulatory monitoring where detailed chemical speciation is unavailable, potentially supporting targeted air-quality management. Uncertainty remains about generalizability, as results are site-specific and the authors note validation across additional sites and conditions is still needed.
- Analysis used four-year dataset (2020-2023) of PM10, PM2.5, NO2, SO2, O3, and meteorological predictors from a single monitoring station.
- Clustering of SHAP-derived predictor-impact profiles identified ten recurrent environmental settings associated with PM10 enhancement, reduction, or transitional behavior.
- Cold-season accumulation setting E7 showed the largest mean impact of 82.9 bcg m-3 with persistent daytime-nighttime occurrence, while E0 was frequent and predominantly nocturnal with mean impact of 50.9 bcg m-3.
Ensemble ML with SHAP clustering achieved R2=0.913 on 2020-2023 regulatory data and identified ten recurrent environmental settings driving PM10, enabling interpretable analysis without chemical speciation.
The rundown
The study analyzed routine regulatory data from one station over 2020-2023, including PM10, PM2.5, NO2, SO2, O3 and meteorology, using ensemble machine-learning models with metaheuristic hyperparameter optimization.
SHAP-derived impact profiles were clustered into ten settings; cold-season accumulation settings drove positive contributions while warm-season and better-mixed settings showed reductions of approximately 26-29 bcg m-3 relative to the model-expected baseline, showing similar concentrations can reflect different pollutant-meteorology configurations.
Sources
- Peer-reviewedScience of The Total Environment2026-08-15
How should this claim be treated?
The debate