Applying traffic camera and deep learning-based image analysis to predict PM_2.5 concentrations.

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Air pollution poses a significant global health burden.
• Existing air quality monitoring networks have limited coverage.
• Cost-effective expansion of monitoring is crucial.

Purpose of the Study:

• To develop a novel method for converting traffic cameras into particulate matter (PM2.5) sensors.
• To expand PM2.5 monitoring coverage without incurring additional infrastructure costs.

Main Methods:

• Collected traffic camera images and hourly PM2.5 data from a reference station.
• Trained a neural network to derive PM2.5 concentration from images.
• Employed a random forest model for hourly PM2.5 prediction.
• Utilized saliency maps and feature importance for model interpretation.

Main Results:

• Achieved high prediction performance for hourly PM2.5 from traffic camera images (RMSE: 0.76 μg/m³, R²: 0.98).
• Model demonstrated robustness across varying weather conditions.
• Saliency maps indicated focus on unobstructed road surfaces; feature importance highlighted image significance.

Conclusions:

• A practical method was developed to repurpose traffic cameras as PM2.5 sensors.
• Deep learning, specifically Resnet architecture, enables expanded PM2.5 monitoring coverage.
• No additional infrastructure is required for this monitoring solution.