Predicting Air Quality Index in Başakşehir, Istanbul with Hybrid AI Models: Unveiling Key Drivers Through Catboost-Based SHAP and Feature Importance Analysis

Abstract

Urban air quality influences public health, ecosystem sustainability and economic productivity. This study focuses on predicting the Air Quality Index (AQI) in Ba & scedil;ak & scedil;ehir, Istanbul. The study proposes a hybrid artificial intelligence (AI) model that amalgamates Categorical Boosting (CatBoost), Shapley Additive Explanations (SHAP) and the feature importance analysis. The dataset encompasses various meteorological parameters, including Tempmax, Tempmin, Temp, Dew, Humidity, Precip, Windspeed, Sea level pressure, Cloud cover, Solar radiation, Solar energy and UV index, in addition to air quality parameters such as PM10, SO2, CO, NO2, NOX, NO and O3. These variables serve as inputs for models like ANN, BAT-ANN,- BBO-ANN, GWO-ANN, HCA-ANN, CatBoost and CNN; the intent is to enhance the accuracy of the AQI prediction. When the combined set of variables were employed as inputs, the most precise results emerged from the CNN model, which yielded an RMSE of 1.43, an AIC of 949.21 and an NSE and R2 of 0.99. The CatBoost model exhibited exceptional performance among the various input combinations, providing the most accurate results for these configurations. Non-parametric statistical Friedman and Nemenyi post-hoc tests were used for multi-model comparison, and it was concluded that there were significant performance differences between the models used according to the p-value values, both in general and based on pairs. While prior studies have explored hybrid AI models for AQI prediction, this study uniquely integrates CatBoost and SHAP for enhanced explainability and model performance evaluation. SHAP analysis provided transparent insights into variable contributions; however, PM10 emerged as the dominant predictor, achieving the highest mutual information score of 6.88. These findings underscore the importance of integrating pollutant and meteorological data. The proposed methodology aligns with global sustainability goals, including SDG 11 (Sustainable Cities and Communities) and SDG 13 (Climate Action).