Interpretable machine learning framework for air quality prediction in Istanbul using Shapley additive explanations (SHAP)

This study develops a season-aware machine-learning (ML) framework to predict hourly concentrations of PM₁₀, PM_2.5 and O₃ across İstanbul. A comprehensive 2021–2023 dataset was compiled from three co-located air-quality and meteorological monitoring stations that typify contrasting source regimes, i.e., a traffic-dominated urban site, a rural background site, and a semi-urban coastal site. Seven ML algorithms, namely eXtreme Gradient Boosting (XGBoost), Extra Trees (ETR), Random Forest (RF), Adaptive Boosting (AdaBoost), Multi-Layer Perceptron (MLP), k-Nearest Neighbors (KNN) and Support Vector Regression (SVR), were utilized to establish a holistic comparison scheme. Hyperparameters were optimized using five-fold cross-validated Bayesian search, and models were evaluated with various performance indicators on season-withheld test sets. In the winter months, ETR achieved a mean R² = 0.93 (RMSE ≈ 10 µg/m³) for PM₁₀ at Bağcılar, while XGBoost yielded R² = 0.88 for O₃ at the same site. Summer predictions were more challenging. PM₁₀ skill in rural Arnavutköy dropped to R² = 0.61 despite strong training fits, highlighting over-fitting risks under complex, non-stationary chemical conditions. By contrast, MLP maintained robust urban performance for PM_2.5 (summer test R² = 0.80) and KNN provided the most stable O₃ prediction in rural areas (R² = 0.74). To enhance interpretability, SHAP (SHapley Additive exPlanations) analysis was applied to the best-performing models, enabling a transparent assessment of how meteorological and co-pollutant inputs shaped predictions at each site. The proposed framework demonstrates that data-driven models can complement traditional air-quality modeling systems by providing station-level insights and interpretable relationships between pollutants and meteorological drivers, supporting air-quality assessment and policy-relevant analyses in rapidly urbanizing regions.