Air quality estimation from sequential surveillance images using a unified CNN-RNN framework

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Traditional air pollution monitoring methods face limitations in cost, resolution, and nighttime applicability.
• Urban environmental management requires accurate, round-the-clock air quality data.

Purpose of the Study:

• To develop a unified deep learning framework for estimating air quality index (AQI) using surveillance images.
• To enable continuous air quality sensing under diverse illumination conditions, including nighttime.

Main Methods:

• A convolutional-recurrent neural network (CNN-RNN) framework was developed to analyze spatial and temporal features from image sequences.
• The model was trained and validated on over 28,000 hourly images from six sites in Kaohsiung, Taiwan.
• The framework was extended to estimate PM2.5 and PM10 levels and adapted to new locations via fine-tuning.

Main Results:

• The unified CNN-RNN model demonstrated superior performance compared to single-image baselines across various sites and times.
• Improved accuracy was observed in higher pollution categories.
• The model successfully estimated air quality under varying illumination, including night and twilight conditions.

Conclusions:

• The developed framework provides a viable solution for round-the-clock, accurate air quality sensing.
• This AI-driven approach enables scalable deployment in existing camera networks, complementing traditional monitoring systems.
• The method offers a cost-effective and high-resolution alternative for urban air quality management.