Machine learning models with explainable AI for PM_2.5 prediction for Urban air quality assessment

Fine particulate matter (PM_2.5) remains a critical air quality issue in rapidly urbanizing Indian cities because of its strong association with respiratory, cardiovascular, environmental, and climatic impacts. This study aimed to develop an accurate and interpretable machine learning framework for predicting PM_2.5 in Nagpur, Maharashtra, India. Approximately 4000 observations, including PM_2.5, NO₂, SO₂, NH₃, O₃, wind speed, wind direction, relative humidity, solar radiation, and temperature, were utilized into 70% training and 30% testing subsets. Six machine learning models decision tree, extra trees, random forest, K-nearest neighbors, XGBoost, and SVR were developed and validated using 10-fold cross validation. The results indicated that the Decision Tree model achieved the highest prediction performance, with R² values of 0.94 and 0.90 for training and testing, respectively with lowest error values, with RMSE of 7.50-8.90 µg/m³, MAE of 5.40-6.60 µg/m³, and MAPE of 10.20-12.80%. SHAP analysis identified NO₂, O₃, RH, and NH₃ as the most influential predictors for processes in PM_2.5variation. The overall model performance followed the order such as DT > ET > RF > KNN > XGBoost > SVR. The developed modeling framework can be applied to PM_2.5 prediction in other regions by incorporating local emission profiles and meteorological dynamics.