A Comparative Study of Conventional and Deep Learning Correction Functions for Low-Cost Particulate Matter Measurements

Abstract

Fine particulate matter smaller than 2.5 micrometers significantly impacts human health, yet its concentrations are governed by complex interactions between emissions and meteorological conditions. This study investigates the seasonal drivers of PM2.5 in the Silesian Voivodeship and assesses the predictive performance of machine learning models, including Light Gradient Boosting Machine, Random Forest, and Support Vector Regression. Meteorological variables, temporal factors, and gaseous pollutants were analyzed to identify key determinants of pollution episodes. During the heating season, elevated residential heating emissions combined with stagnant atmospheric conditions dominate PM2.5 variability, while in the non-heating season, atmospheric dynamics and seasonal timing play a stronger role. Light Gradient Boosting Machine consistently achieved the highest predictive accuracy, capturing nonlinear interactions and providing interpretable insights into feature importance. These results highlight the potential of machine learning approaches for accurate air quality modeling and inform strategies for pollution mitigation and public health protection.