Spatiotemporal analysis and machine learning-based prediction of air quality in Indian urban cities

Air pollution, more specifically Particulate Matter (PM2.5 - particulate matter with diameter less than 2.5 micrometers), threatens the public health most critically in urban Indian cities, and Delhi, among them, presents the most acute challenge. This study predicts the concentrations of PM2.5 using machine learning models using data ranging from 2010 to 2023 and assessing model fit via R², RMSE, MAE, and MAPE metrics. Models tested: Random Forest, Gradient Boosting, AdaBoost, Histogram-Based Gradient Boosting, XGBoost. The Random Forest model is extremely effective for the training set (R² = 0.99) but shows the highest degree of overfitting, with R² of 0.35 for the test set. Gradient Boosting has a more balanced result, with R² 0.54 and 0.48, respectively on the training and test set, as well as fewer errors (RMSE: 56.46, MAE: 39.60, MAPE: 0.50). Hence, it is a good predictor. AdaBoost performs the worst with an R² of 0.28 on the test set and the highest errors in terms of RMSE: 66.86, MAE: 52.34, MAPE: 0.94. Histogram Gradient Boosting and XGBoost: both of these models yield an average accuracy value, but the Gradient Boosting model is still a tad better than the former ones in terms of RMSE and MAE. Thus, Gradient Boosting happens to be the most accurate model in light of generalization as well as accuracy for the prediction of the concentration of PM2.5. These results will be highly beneficial to policymakers to adopt machine learning-based air quality forecasting for better environmental management and the protection of public health.