Scientific Publications

NextAIRE Researchers Investigate Graph Neural Networks for Urban Air Quality Estimation

Researchers associated with the NextAIRE consortium have contributed to a new scientific preprint examining the effectiveness of graph neural networks (GNNs) in estimating air pollution at urban locations without dedicated monitoring stations.

The study, “How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation,” was published on ChemRxiv on 2 September 2026.

The research brings together experts from several institutions, including NextAIRE partners the Lisbon Council, the Institute for Medical Research and Occupational Health (IMROH) and the Institute for Anthropological Research (INANTRO).

Exploring Graph Neural Networks for Air Quality Monitoring

Accurate air quality information is essential for understanding urban pollution patterns and supporting evidence-based environmental decisions.

However, air quality monitoring stations cannot cover every location within a city, creating challenges for estimating pollution levels in unmonitored areas.

Graph neural networks offer a promising approach by modelling relationships between monitoring stations, traffic information and environmental conditions.

The researchers developed and evaluated a graph attention network that incorporates both air quality monitoring stations and traffic counters, while accounting for wind-related information.

The study used data collected between 2018 and 2021 from five air quality monitoring stations and 16 traffic counters in Graz, Austria.

Key Research Findings

An important objective of the study was to determine whether the performance of graph neural networks results from their advanced architecture or simply from access to additional information.

To investigate this, the researchers compared the graph-based approach with six alternative statistical and machine learning models, ensuring that all methods had access to the same information.

The findings revealed that:

  • Access to neighbouring monitoring stations’ measurements was a major factor influencing prediction performance.
  • Graph neural networks did not demonstrate a meaningful performance advantage over simpler approaches when evaluated using equivalent information.
  • Reliable model evaluation is particularly important when predicting air quality at locations without direct measurements.
  • Uncertainty estimation and transparent evaluation procedures are essential for assessing the reliability of AI-based predictions.

These findings highlight the importance of evaluating advanced AI methods against appropriate alternatives before applying them to environmental monitoring challenges.

Advancing Reliable and Trustworthy AI for Air Quality Research

The study provides valuable insights into the strengths and limitations of graph neural networks for urban air quality estimation.

Rather than assuming that increasingly complex AI models always deliver better results, the research emphasises the importance of fair comparisons, methodological transparency and robust validation.

This approach closely aligns with NextAIRE’s objectives of advancing reliable AI-driven environmental monitoring, strengthening interdisciplinary research and supporting innovative approaches to air quality analysis.

The study is currently available as a research preprint on ChemRxiv and has not yet undergone peer review.

Read the Full Research Preprint

Title: How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation

Authors: Valentino Petrić, Dejan Strbad, Nikolina Račić, Francesco Mureddu, Andreas Schopper, Benedikt Tschofenig, Michael Haberl and Mario Lovrić

Platform: ChemRxiv

Date posted: 2 September 2026

Full preprint: https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008206/v1