In brief
Air quality at one monitoring station is connected to conditions elsewhere. This work learns both nearby and long-range relationships across the Spanish monitoring network.
Contribution
A graph architecture that combines hierarchical message passing, spectral information from the adjacency matrix and adaptive fusion of pollutant and meteorological variables.
Key finding
The approach outperformed the evaluated deep-learning baselines across the atmospheric pollutants considered.
Models and data
Distinctive research elements
Models and methods
- Hierarchical message passing
- SVD graph representation
Data and evaluation
- Spanish air-quality monitoring network
- Meteorological variables
- 2010–2020 observations
Author-written overview
Research overview
Air-quality conditions at one monitoring station depend on both nearby observations and processes acting across longer distances. This work models those relationships using pollutant and meteorological records from a Spanish monitoring network covering 2010–2020. The architecture combines hierarchical message passing with spectral information obtained through singular value decomposition of the adjacency matrix, allowing local exchange and global graph structure to contribute to the forecast. A multilayer perceptron performs adaptive feature fusion and a trainable aggregation mechanism combines messages. Comparisons with contemporary deep-learning baselines show better results across the evaluated pollutants, supporting the combination of graph spectral information and spatiotemporal learning for inter-city forecasting.
Cite this work
BibTeX
@article{vicent2026inter,
title = {Spatio-temporal graph neural network for inter-city air quality forecasting},
author = {Jose F. Vicent and Manuel Curado and Marc Semper},
journal = {International Journal of Environmental Science and Technology},
year = {2026},
volume = {23},
number = {1},
eid = {63},
doi = {10.1007/s13762-025-06850-2}
}Always check the publisher record for final volume, issue and page information before citing.