Real-time image-based air quality estimation by deep learning neural networks.

Area of Science:

• Environmental Science
• Computer Science
• Public Health

Background:

• Air quality monitoring is crucial for public health and environmental equity.
• Current methods often rely on expensive instrumentation, limiting accessibility.
• There is a need for efficient, inexpensive, and scalable air quality assessment tools.

Purpose of the Study:

• To develop and validate an image-based deep learning model for estimating air quality.
• To assess the model's accuracy in simultaneously predicting multiple pollutants (PM2.5, PM10, AQI).
• To provide a cost-effective alternative to traditional air quality monitoring systems.

Main Methods:

• A deep learning model (CNN-RC) integrating a convolutional neural network (CNN) and a regression classifier (RC) was developed.
• The model utilizes image features (current, baseline, HSV statistics) for air quality estimation.
• Trained and tested on 3549 hourly air quality datasets from Taiwan, including images and pollutant levels.

Main Results:

• The model achieved high estimation accuracy (R²): 76% for PM2.5, 84% for PM10, and 76% for AQI using daytime images.
• Nighttime images showed improved accuracy for PM2.5 (83%) and similar accuracy for PM10 (84%) and AQI (74%).
• Demonstrated capability for simultaneous, accurate multi-pollutant estimation from single images.

Conclusions:

• The proposed image-based deep learning model provides a promising solution for rapid and reliable air quality estimation.
• This scalable approach can help bridge the gap in air quality data availability globally.
• Offers a cost-effective method for assessing air quality and supporting pollution control efforts.