Urban air quality forecasting requires transparent AI systems to support policymaking and public trust. This study presents an explainable AI (XAI) framework integrating geostationary satellite imagery (GEMS), IoT sensor networks (PM2.5, NO2), and real-time traffic data for high-resolution pollution forecasting. Our hybrid model combines temporal convolutional networks with attention mechanisms, achieving 92.4% accuracy in 24-hour PM2.5 predictions across 15 megacities. SHAP analysis reveals traffic flow contributes 38% to urban NO2 variability, while satellite thermal bands account for 27% of ozone forecasts. The system reduces prediction errors by 40% compared to conventional LSTM models, with integrated Grad-CAM visualizations explaining localized pollution hotspots. Field tests demonstrate 89% accuracy in identifying industrial emission sources using multi-modal data fusion, enabling targeted mitigation strategies. This work bridges the gap between complex AI predictions and actionable urban planning insights through model-agnostic interpretability techniques [1].