Detecting Elevated Air Pollution Levels by Monitoring Web Search Queries: Algorithm Development and Validation
Chen Lin1, Safoora Yousefi1, Elvis Kahoro2
1Department of Computer Science, Emory University, Atlanta, GA, United States.
JMIR Formative Research
|December 19, 2022
Summary
This study shows that using web search data alongside meteorological data can improve real-time air quality nowcasting for ozone, nitrogen dioxide, and fine particulate matter (PM2.5). This approach offers a more accessible way to monitor pollution levels and their public health impacts.
Area of Science:
- Environmental Science
- Computer Science
- Public Health
Background:
- Real-time air pollution monitoring is crucial for public health and environmental surveillance.
- Traditional air quality models rely on ground-based monitors and meteorological data, limiting accessibility and public health impact prediction.
- Artificial neural networks have increased in air pollution research, but models based on physical measurements have limitations.
Purpose of the Study:
- To develop and validate models for nowcasting observed air pollution levels using publicly available web search data.
- To assess the effectiveness of machine learning models, including deep learning, in predicting air quality using meteorological and web search data.
- To explore novel deep learning architectures for improved air pollution nowcasting.
Main Methods:
- Developed machine learning models (supervised classification and deep learning) using meteorological and Google Trends data.
- Validated models for predicting ozone (O3), nitrogen dioxide (NO2), and fine particulate matter (PM2.5) in 10 major US metropolitan areas.
- Explored Long Short-Term Memory (LSTM) variations, proposing a novel dictionary learner-LSTM model for sequential pattern analysis.
Main Results:
- The top-performing model, a deep neural sequence LSTM, achieved high accuracy in detecting elevated pollution levels: 0.82 for O3, 0.74 for NO2, and 0.85 for PM2.5.
- Incorporating web search data significantly improved model accuracy compared to using meteorological data alone.
- The F1-scores for detecting elevated pollution were 0.51 for O3, 0.41 for NO2, and 0.27 for PM2.5.
Conclusions:
- Integrating web search data with meteorological data enhances the nowcasting performance for key air pollutants.
- This approach demonstrates a promising new method for tracking global physical phenomena using readily available web search data.
- The findings suggest broader applications for web search data in environmental monitoring and public health surveillance.
