Urban Air-Quality Estimation Using Visual Cues and a Deep Convolutional Neural Network in Bengaluru (Bangalore),
Alon Feldman1, Shai Kendler2,3, Julian Marshall4
1Department of Mathematics, Technion-Israel Institute of Technology, Haifa 3200003, Israel.
Environmental Science & Technology
|December 17, 2023
Summary
This study introduces a deep learning method using dashboard camera videos to estimate air pollution levels. The approach effectively infers pollutant concentrations, offering a cost-effective solution for mobile monitoring.
Area of Science:
- Environmental Science
- Computer Science
- Data Science
Background:
- Mobile monitoring offers detailed air pollution data but faces limitations due to resource constraints.
- Static images have been explored as surrogates for air pollution assessment, but with limitations.
- Developing cost-effective, scalable methods for air quality monitoring is crucial.
Purpose of the Study:
- To develop and evaluate deep learning methodologies for inferring on-road pollutant concentrations using video data from dashboard cameras.
- To assess the performance of a convolutional neural network (CNN) approach against other machine learning techniques for air pollution estimation.
- To enable real-time analysis of air quality using readily available video technology.
Main Methods:
- Collected 50 hours of on-road air pollution measurements (black carbon, particle number concentration, PM2.5, CO2) in Bengaluru, India.
- Utilized deep learning, specifically a regression CNN, to analyze video frames, identifying objects and motion (segmentation, optical flow).
- Compared CNN performance against traditional linear regression and other machine learning models.
Main Results:
- The CNN approach significantly outperformed other machine learning techniques and conventional analyses in predicting pollutant concentrations.
- The CO2 prediction model demonstrated high accuracy, with a normalized root-mean-square error between 10-13.7%.
- The study successfully used video and object motion, rather than static images, for pollution inference.
Conclusions:
- Deep learning analysis of dashboard camera videos provides a novel and effective method for mobile air pollution monitoring.
- This video-based approach offers a rapid, real-time analysis capability, overcoming limitations of traditional methods.
- The methodology is adaptable to other mobile monitoring campaigns, requiring only inexpensive dashboard cameras.
