Development and interpretation of PM2.5 estimation model for the Seoul Metropolitan Area using machine learning and explainable AI

Abstract

PM_2.5 is emitted and formed in the atmosphere through various factors, posing significant health risks to humans. Therefore, accurately estimating PM_2.5 concentrations and analyzing the contributions of individual factors are crucial. A Deep Neural Network (DNN) model was developed for PM_2.5 estimation in the Seoul Metropolitan Area in South Korea, while some machine learning models—Random Forest and Extreme Gradient Boosting—were also built for performance comparison. Among these, the DNN model demonstrated the best performance, with an R² of 0.95, MSE of 12.14, and MAE of 2.6. Based on this, Explainable Artificial Intelligence (XAI) techniques, including Vanilla Gradient and Shapley Additive Explanation (SHAP), were applied to interpret the PM_2.5 estimation model and analyze the contribution of each factor. The contribution analysis for the Seoul Metropolitan Area revealed that NO₃⁻ and NH₄⁺ had the highest contributions to PM_2.5 formation, indicating that secondary formation mechanisms play a dominant role. Furthermore, at high concentrations, the contributions of NO₃⁻, NH₄⁺, and SO₄²⁻ were the highest, and the contributions of metal components and PM₁₀ were higher than the average. In particular, it was observed that NH₄⁺ and K showed a positive correlation with PM_2.5 formation. Future research will focus on refining the model through clustering-based approaches and other enhancements, aiming to deepen the understanding of PM_2.5 formation patterns and provide meaningful insights for policymaking.