Explainable hybrid deep learning framework for air quality index prediction in Indian cities

Abstract

Air pollution constitutes one of the most pressing environmental and public health crises of the twenty-first century, disproportionately affecting rapidly urbanising nations such as India. Accurate prediction of the Air Quality Index (AQI) is essential for enabling timely hospital preparedness, school-closure decisions, and evidence-based policy interventions. AQI prediction remains inherently challenging owing to widespread missing data due to sensor failures (up to 54% missingness in certain pollutant columns), pronounced non-stationarity, nonlinear inter-pollutant interactions, and the limited interpretability of existing predictive models. This paper presents a four-stage Explainable Hybrid Deep Learning Framework validated on 435,742 CPCB India records (1990–2015) to address these challenges. The XAI-Enhanced Ensemble achieves RMSE = 0.3817, MAE = 0.2773, and R2= 0.9994 on the held-out test set a 54.7% RMSE and 58.7% MAE reduction over the standalone LSTM baseline. SHAP-guided feature selection reduces input dimensionality by 62% (from 26 to 10 features) while yielding an additional 38% RMSE improvement over the standard ensemble, demonstrating that interpretability and accuracy are complementary rather than competing objectives. A four-stage Explainable Ensemble Framework integrating noise-augmented temporal imputation, LSTM (sequence length L = 14), XGBoost, and SHAP-guided feature selection achieves RMSE = 0.3817 and R2= 0.9994 on 435,742 real-world CPCB India records. A noise-augmented linear interpolation pipeline (σₙ = 0.02·σ_col) handles up to 54% missingness without requiring adversarial training, with formal variance-preservation guarantee. SHAP-guided selection of the top 10 features (62% dimensionality reduction) yields a 38% additional RMSE reduction over the standard ensemble, demonstrating that interpretability and accuracy. Paired t-test and Wilcoxon signed-rank tests confirm statistical significance. A four-stage Explainable Ensemble Framework integrating noise-augmented temporal imputation, LSTM (sequence length L = 14), XGBoost, and SHAP-guided feature selection achieves RMSE = 0.3817 and R2= 0.9994 on 435,742 real-world CPCB India records. A noise-augmented linear interpolation pipeline (σₙ = 0.02·σ_col) handles up to 54% missingness without requiring adversarial training, with formal variance-preservation guarantee. SHAP-guided selection of the top 10 features (62% dimensionality reduction) yields a 38% additional RMSE reduction over the standard ensemble, demonstrating that interpretability and accuracy. Paired t-test and Wilcoxon signed-rank tests confirm statistical significance.