Machine learning –driven reclassification of air quality zones and PM2.5 prediction: A case study in a subtropical region

Air quality prediction is crucial for effective environmental management. However, traditional approaches often rely on administrative zoning schemes that fail to capture fine-scale variations in pollution patterns. This gap arises because administrative boundaries misrepresent the dynamic nature of PM2.5, reducing the accuracy, spatial resolution, and policy relevance of machine learning-based prediction models. To address this, a spatially adaptive prediction framework was developed, integrating data-driven reclassification of air quality zones with machine learning techniques. A comprehensive five-year dataset (2018–2022) from 74 ground-based monitoring stations across Taiwan was analyzed using Radial Basis Function (RBF) interpolation and Hierarchical Agglomerative Clustering (HAC). This analysis delineated 14 empirically derived air quality zones based on spatiotemporal similarities in PM2.5 concentration patterns. These refined zones were incorporated into four machine learning models to evaluate the impact of spatial zoning on prediction performance. The results indicated that the reclassified zones substantially enhanced model accuracy, with R 2 values reaching 0.85 and reducing RMSE by over 40 % in pollution-intensive regions. This framework provides a scalable and transferable approach for high-resolution PM2.5 prediction, improving exposure assessment and enabling targeted air quality interventions. The findings contribute to spatial air quality modeling by demonstrating that empirically derived zoning can enhance PM2.5 prediction reliability and support the development of data-driven air quality management strategies.