An Explainable Transfer Learning Framework for PM2.5 Forecasting in Data-Scarce Urban Environments Using Cross-Regional Atmospheric Knowledge Transfer

Fine particulate matter (PM2.5) remains one of the most critical environmental and public-health challenges worldwide, particularly in developing regions where air-quality monitoring networks are sparse and long-term observational records are often unavailable. Although recent advances in machine learning have substantially improved PM2.5 forecasting, most existing models depend on large volumes of locally labelled data and exhibit poor generalization when transferred to new cities with different climatic and pollution characteristics. Furthermore, current transfer-learning approaches primarily emphasize predictive accuracy, while providing limited understanding of whether the transferred atmospheric knowledge remains physically meaningful and interpretable across regions. To address these limitations, this study proposes an explainable transfer learning framework for PM2.5 forecasting in data-scarce urban environments by transferring atmospheric knowledge learned from data-rich cities to regions with limited monitoring data. The framework combines deep neural networks with domain adaptation to capture transferable relationships between PM2.5 concentrations and key meteorological variables, including temperature, relative humidity, wind speed, precipitation, and atmospheric pressure. Unlike existing studies that mainly transfer model parameters or latent features, the proposed framework explicitly integrates cross-regional atmospheric knowledge transfer with Explainable Artificial Intelligence (XAI) to reveal how meteorological drivers influence predictions after knowledge transfer, thereby enabling both accurate forecasting and scientific interpretation of the transferred representations. This dual emphasis on transferability and interpretability provides greater confidence in applying the model to previously unseen urban environments and facilitates a better understanding of region-specific pollution dynamics. The proposed framework is evaluated across multiple cities representing diverse climatic and air-pollution regimes to assess its robustness, transferability, and generalization capability under limited-data conditions. By reducing dependence on extensive local monitoring while maintaining transparent, interpretable predictions, this work offers a scalable framework for next-generation atmospheric intelligence systems that can support air-quality management, public-health protection, and sustainable urban development, particularly in regions where environmental monitoring infrastructure remains limited.