Machine Learning, Deep Learning, and Explainable AI for Multi-Source AIr Pollution Forecasting: A Review

Air pollution remains one of the most pressing environmental and public health challenges of the current decade, motivating sustained research into computational methods for forecasting pollutant concentrations. This review synthesizes recent literature on machine learning (ML), deep learning (DL), hybrid ensemble architectures, multi-source data integration, and explainable artificial intelligence (XAI) as applied to air quality forecasting. Tree-based ensemble methods such as XGBoost, LightGBM, and CatBoost, alongside recurrent architectures such as Long Short-Term Memory (LSTM) networks and their convolutional and attention-augmented variants, dominate the recent literature, frequently combined into hybrid or stacked frameworks to balance short-term responsiveness with long-term temporal dependency modeling. A growing body of work integrates heterogeneous data streams, including ground-based sensor networks, meteorological records, traffic and emission inventories, and satellite-derived aerosol and trace-gas products, to improve spatial and temporal coverage. In parallel, explainability techniques, most notably SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), are increasingly incorporated to address the opacity of high-performing black-box models and to support policy-relevant decision-making. This review consolidates findings across these strands, identifies methodological gaps relating to model transparency, heterogeneous data fusion, and generalizability across geographic contexts, and outlines directions for future hybrid explainable forecasting frameworks.