A Dual-Branch Explainable AI Framework Combining Machine and Deep Learning for Air Quality Forecasting in Indian Cities

Pratik Chakraborty, Shanthi P B

Abstract A precise Air Quality Index (AQI) forecasting is essential for ecosystem monitoring and societal health management. This study proposes a dual-branch explainable AI framework that brings together machine learning models (XGBoost, LightGBM,CatBoost, Linear Regression) and deep learning models (LSTM, GRU, TCN, Transformer) for AQI prediction across major Indian cities. Comprehensive preprocessing, including temporal feature extraction and rolling mean engineering, was applied to ensure high-quality and reliable inputs. The GRU model achieved the best performance with an R2 score of 0.9886 and an RMSE of 0.0032. A multi-level XAI analysis was incorporated, using Feature Importance for machine learning models and Integrated Gradients, Perturbation Sensitivity, Feature Occlusion, and Temporal Receptive Field Profiling for deep learning models. The results consistently identified AQI_7d_avg and AQI_30d_avg as the most influential predictors. The proposed framework combines strong predictive accuracy with transparent interpretability, supporting data-driven environmental policymaking and sustainable air pollution control system across Indian cities.