AQIFormer: A Transformer-Based Multi-View Architecture for Cross-City Air Quality Classification

Researchers have developed AQIFormer, a transformer-based AI system that estimates air quality from traffic camera imagery combined with weather data. The model achieves 89.96% accuracy on training data and maintains strong cross-city generalization with 81.67% accuracy on independent Indian datasets, significantly outperforming existing methods.

AQIFormer addresses a genuine infrastructure challenge in environmental monitoring by replacing or supplementing expensive sensor networks with computer vision technology. Traditional air quality monitoring relies on distributed sensor stations that are capital-intensive and difficult to scale across developing regions, creating blind spots in pollution tracking. This research demonstrates that visual atmospheric characteristics captured in standard traffic cameras, when processed through advanced transformer architectures, can reliably estimate pollutant levels.

The technical innovation centers on multi-view integration—combining front and rear dashcam perspectives with meteorological parameters through attention mechanisms sensitive to weather conditions. This approach reflects a broader trend in environmental AI where researchers leverage existing infrastructure (traffic cameras are ubiquitous in urban areas) to generate new data streams. The 14.96% improvement over prior methods is substantial, but the cross-city generalization results are particularly significant: achieving 81.67% accuracy in Nagpur, India with only 8.29% degradation demonstrates the model's robustness beyond training environments.

For environmental monitoring stakeholders, this technology presents clear practical value. Cities could retrofit existing traffic surveillance systems to provide continuous air quality estimates without additional sensor deployments, reducing costs substantially. The few-shot adaptation capability suggests rapid deployment in new cities requires minimal localized training data. However, the technology's impact remains confined to specific use cases—visual estimation cannot replace chemical sensors for regulatory compliance or detailed pollutant identification, limiting commercialization potential. Developers building environmental monitoring platforms could integrate this approach as a supplementary data source, while municipalities might pilot the system for localized pollution hotspot detection and public health warnings.

• →AQIFormer uses transformer architecture and dual-view traffic imagery to estimate air quality with 89.96% accuracy, outperforming prior methods by 14.96%.
• →The model maintains exceptional cross-city generalization with only 8.29% performance degradation when applied to independent datasets in India.
• →Integration of weather-aware attention mechanisms enables the system to account for meteorological factors affecting air quality estimation.
• →Few-shot adaptation capabilities allow rapid deployment in new cities with minimal additional training data.
• →The approach leverages existing traffic camera infrastructure to supplement or replace expensive traditional air quality sensor networks.