DOI: 10.21541/apjess.1968813 ISSN: 2147-4575

Machine Learning-Based Prediction of PM2.5 Levels in OECD Countries: An Integrated AutoML, SHAP, and Scenario Analysis Approach

Beyza Yıldız, Enes Furkan Erkan
Show PDF Cite
Air pollution remains a major challenge for environmental sustainability and public health. PM2.5, defined as fine particulate matter, is a critical air pollution indicator because its small particle size can adversely affect the respiratory and cardiovascular systems. Monitoring and predicting PM2.5 levels are essential for air quality policy, risk reduction, and health-based decision-making. This study predicts PM2.5 levels in OECD countries using environmental, socioeconomic, urban, and transportation-related indicators. A country-year-level dataset comprising 264 observations from 24 OECD countries during 2013–2023 was constructed. PM2.5 was used as the dependent variable, while built-up area ratio, green coverage ratio, industrial indicators, population density, urbanization rate, economic density, environmental protection, environmental regulation, internal combustion engine vehicles, and electric vehicles were included as independent variables. Given the temporal structure, observations from 2013–2021 were used for model development, while observations from 2022–2023 constituted the holdout set. An AutoML approach was employed to compare 18 regression algorithms using performance metrics. The Extra Trees Regressor achieved the best cross-validated training performance, with an R² of 0.914, and obtained an R² of 0.971 on the temporal holdout set. SHAP analysis identified electric vehicles, population density, environmental protection, green coverage ratio, and internal combustion engine vehicles as the most influential predictors. Scenario analysis showed that increasing electric vehicles and green coverage while reducing internal combustion engine vehicles and built-up areas decreased predicted PM2.5 level by 5.42%. Overall, PM2.5 should be evaluated through the combined effects of transportation, demographic, environmental, and urban structural indicators rather than a single variable.

More from our Archive