Abstract
Traditional air quality monitoring relies on sparse fixed stations, failing to capture hyper-local variations in on-road pollutants. This study presents an optimized strategy for predicting street-level air pollution by integrating mobile monitoring data with deep learning analysis of Street View Images (SVIs). By analyzing over 382,000 images from Guangzhou, China, researchers identified that a Random Forest model utilizing an averaging strategy across four camera angles within a 100m radius yields the highest predictive accuracy.
Introduction
On-road air pollution, including nitrogen oxides (NOx) and particulate matter (PM), fluctuates significantly over short distances due to urban morphology, traffic patterns, and vegetation. While mobile monitoring offers higher resolution than fixed stations, it remains costly to scale. Street View Images (SVIs) provide a ubiquitous visual proxy for the urban environment, yet a standardized methodology for their use in pollution modeling has been lacking. This research establishes a "gold standard" for SVI-based prediction, evaluating machine learning models, sampling radii, and image quality.
Methodology
The researchers collected nearly 5 million data points for NO, NO₂, PM₂.₅, and PM₁₀ using sensors mounted on 314 taxis in Guangzhou. Simultaneously, 382,755 SVIs were retrieved via the Baidu API.
The analytical pipeline involved several critical steps:
- Semantic Segmentation: Utilizing Mask2Former, a state-of-the-art Transformer-based model, to extract 19 environmental features (e.g., vehicles, trees, buildings) from the images.
- Spatial Analysis: Testing various buffer radii (100m to 500m) to determine the optimal area of influence for pollutant concentration.
- Viewpoint Strategy: Comparing individual camera angles (0°, 90°, 180°, 270°) against an averaging strategy that combines all four directions to capture the full environmental context.
- Model Benchmarking: Comparing Random Forest (RF), XGBoost, and Neural Networks against traditional Land-Use Regression (LUR) models.
Key Findings
The Optimal Spatial Scale
Contrary to expectations that more data is always better, the study found that a 100m radius buffer outperformed larger scales (up to 500m). Expanding the radius introduced "noise"—features irrelevant to the immediate on-road environment—which diluted the model's predictive power. This suggests that on-road pollution is primarily driven by hyper-local factors.
The 360° Perspective
While road-facing images (0° and 180°) effectively capture traffic-related features, the averaging strategy across all four cardinal directions proved most robust. This approach mitigates bias and provides a comprehensive view of the surrounding urban canyon, including sidewalks and buildings that influence pollutant dispersion and human activity levels.
Model Performance and Non-Linearity
The Random Forest model emerged as the superior algorithm, followed by XGBoost and Neural Networks. All machine learning models significantly outperformed the linear LUR model. This highlights the complex, non-linear relationship between urban visual features and air quality. The optimal strategy achieved absolute errors of less than 2.5 μg/m³ or ppb for most pollutants.
The Impact of Image Quality
Data cleaning proved critical to accuracy. Automated filtering of low-quality images (blurry or overexposed), which accounted for roughly 16% of the dataset, improved prediction accuracy by 10% to 24%. Poor lighting often leads models to overestimate road features while underestimating human-activity indicators, causing significant estimation errors.
Conclusion
This research provides a scalable, cost-effective framework for hyper-local air quality mapping. By optimizing the spatial radius and viewpoint strategy, urban planners can leverage existing SVI databases to monitor pollution with unprecedented detail. Future work will focus on global scalability and integrating real-time visual data to account for dynamic urban changes like construction or seasonal vegetation shifts.
Summary updated