Machine learning-based approaches to forecast particulate matter in the Istanbul metropolitan area

Abstract

Air pollution poses a major global health challenge, with particulate matter (PM) linked to millions of premature deaths each year. This study forecasts PM concentrations in Istanbul's Kartal district using bi-daily observations collected throughout 2022. Four supervised machine learning (ML) models, including support vector machines, random forests (RFs), artificial neural networks, and K-nearest neighbors, were applied using surface meteorological variables and radiosonde-derived inversion parameters. The RF model achieved the highest predictive accuracy, with R² values of 0.64 for PM₁₀ and 0.70 for PM_2.5, along with the lowest mean squared error. The study incorporates key enhancements, including the integration of vertical inversion metrics with surface pollutant data, the use of autocorrelation analysis to justify lagged features, and statistical evaluation of model differences using paired t-tests. Feature importance analysis showed that inversion thickness and lagged PM levels improved forecasts, highlighting the value of upper-air dynamics and temporal persistence. The aim of this study is to systematically evaluate multiple ML algorithms for PM forecasting at a single urban site. The findings provide transparent, site-specific methodological insights that highlight the role of upper-air dynamics and temporal persistence, offering practical implications for similar urban environments and guiding future multi-site applications.