Fine-Grained Air Quality Inference via Multi-Channel Attention Model

Qilong Han, Dan Lu, Rui ChenIJCAI 2021 · DOI

Abstract

Fine-grained air quality inference—estimating pollutant levels at locations without monitoring stations—is a critical task for urban health and policy. However, existing models often fail to distinguish between static geographic features and dynamic atmospheric conditions. This paper presents the Multi-Channel Attention Model (MCAM), a dual-channel graph approach that explicitly models static geo-context and dynamic physics-informed atmospheric dispersion. Evaluated on real-world datasets from Beijing and London, MCAM demonstrates significant improvements in inference accuracy over state-of-the-art baselines.

Introduction

Air quality monitoring stations are expensive and sparse, creating a need for high-resolution data to inform public health decisions. The challenge of fine-grained air quality inference lies in the complex interplay between static factors (e.g., road networks, points-of-interest) and dynamic factors (e.g., wind speed, pollutant travel time). Existing data-driven models often treat these spatial correlations as uniform, failing to account for the physical laws governing air movement.

The MCAM Architecture

The proposed Multi-Channel Attention Model (MCAM) addresses these challenges by separating spatial relationships into two distinct channels before fusing them for final inference.

The MCAM architecture consists of a Static channel for geographic features, a Dynamic channel for temporal/atmospheric trends, and a Fusion module.

Static Graph Channel

This channel models fixed geographic relationships using an adapted bilateral filtering technique. By calculating weights based on "geo-context" similarity (road networks and POIs) rather than simple Euclidean distance, the model ensures that stations with similar surroundings have a greater influence on the target location.

Dynamic Graph Channel

To capture changing atmospheric conditions, this channel utilizes Long Short-Term Memory (LSTM) networks to model temporal trends. It incorporates a physics-informed attention mechanism based on:

  • Atmospheric Dispersion Theory: Using wind velocity and relative angles to model pollutant drift.
  • Hysteretic Effects: Accounting for the time-lag between a pollutant's presence at a source station and its arrival at the target location.

Multi-Channel Fusion

A specialized Graph Convolutional Network (GCN) fuses the outputs from both channels with non-graph features to produce the final air quality estimate.

Experimental Evaluation

The model was evaluated on two major real-world benchmark datasets from Beijing (35 stations) and London (13 stations). The evaluation used a "leave-some-out" strategy to validate inference accuracy against ground truth at unseen locations.

Key Results and Discussion

MCAM consistently outperformed representative baselines, including K-Nearest Neighbors (KNN), Linear Interpolation (LI), and the previous state-of-the-art model, ADAIN.

Performance comparison of MCAM against baselines across PM2.5, PM10, and NO2 pollutants.

As shown in the results, MCAM achieved a significant accuracy boost. On the Beijing dataset, it improved Mean Absolute Error (MAE) by up to 17.5% and Root Mean Squared Error (RMSE) by 16.5% compared to ADAIN. The model's performance remained robust across different pollutants (PM2.5, PM10, and NO2) and geographic settings, suggesting that its physics-guided approach captures fundamental principles of air movement.

Ablation studies further highlighted that while both channels contribute, the dynamic graph channel is the most critical component, reflecting the fluid nature of air pollution. Additionally, accounting for the hysteretic nature of pollutant transport provided a measurable boost in prediction accuracy.

Conclusion

MCAM provides a powerful framework for fine-grained air quality inference by explicitly modeling the static and dynamic factors influencing pollution. By integrating physics-informed attention and bilateral filtering, the model bridges the gap between purely data-driven approaches and physical reality. Future work may involve integrating higher-resolution auxiliary data, such as satellite imagery, to further refine inference in data-sparse regions.

Summary updated

BibTeX
@inproceedings{han2021fine,
  author = {Qilong Han and Dan Lu and Rui Chen},
  title = {{Fine-Grained Air Quality Inference via Multi-Channel Attention Model}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/346},
}
Figure