Predicting On-Road Air Pollution Coupling Street View Images and Machine Learning: A Quantitative Analysis of the...

Area of Science:

• Environmental Science
• Remote Sensing
• Data Science

Background:

• Predicting local air pollution using mobile monitoring and street view images (SVIs) shows potential.
• Existing methods lack reliable references to quantify errors from algorithms, sampling, and image quality.

Purpose of the Study:

• To quantify the impact of algorithms, sampling strategies, and image quality on air pollution prediction using SVIs.
• To identify optimal strategies for integrating mobile monitoring data and SVIs for accurate air quality modeling.

Main Methods:

• Deployed 314 taxis for real-time monitoring of NO, NO2, PM2.5, and PM10.
• Extracted features from ~382,000 SVIs at multiple angles and buffer radii (100-500m).
• Compared machine learning algorithms (Random Forest, XGBoost, Neural Network) against a land-use regression (LUR) model.

Main Results:

• Machine learning methods generally outperformed linear LUR models.
• An averaging strategy effectively reduced bias from insufficient feature capture.
• Optimal sampling involved integrating multiple viewing angles within a 100m buffer, yielding errors < 2.5 μg/m³ or ppb.
• Image quality issues (over/underexposure, blur) led to misidentification of features.

Conclusions:

• The study provides valuable insights into optimizing SVI-based air quality models.
• Findings support the development of more accurate and reliable environmental monitoring systems.
• Best practices include using multiple viewing angles and appropriate buffer sizes for feature extraction.