Convolutional Neural Networks Facilitate Process Understanding of Megacity Ozone Temporal Variability

Area of Science:

• Atmospheric Chemistry
• Environmental Science
• Artificial Intelligence in Environmental Monitoring

Background:

• Ozone pollution is heavily influenced by meteorological conditions like temperature, pressure, wind, and humidity.
• Existing empirical models often overlook high-altitude and regional meteorological influences on ozone variability.
• Megacity air quality management requires understanding complex, spatiotemporal meteorological effects on ozone dynamics.

Purpose of the Study:

• To investigate the influence of three-dimensional meteorological fields on ozone dynamics in Shenzhen, China, using convolutional neural networks (CNNs).
• To quantify the contribution of various meteorological factors to daily, seasonal, and interannual ozone variability.
• To develop an interpretable AI framework for attributing ozone fluctuations to nonlinear meteorological effects.

Main Methods:

• Application of convolutional neural networks (CNNs), typically used for image recognition, to analyze meteorological data and ozone levels.
• Development of an optimized CNN model covering a 13° × 13° spatial domain to capture regional influences.
• Utilizing model interpretation techniques to identify key meteorological drivers of ozone variability.

Main Results:

• The CNN model explains over 70% of daily ozone variability, outperforming traditional empirical models.
• 2-m temperature (16%) and humidity (15%) are identified as primary drivers of daily ozone fluctuations.
• Regional wind fields contribute up to 40% to ozone changes during specific episodes.
• CNNs attribute -5-6 μg·m^-3 of interannual ozone variability to weather anomalies.

Conclusions:

• Interpretable CNNs offer a powerful tool for understanding nonlinear meteorological impacts on ozone across spatiotemporal scales.
• The findings provide essential process-based insights for effective air quality management in megacities under changing climate conditions.
• This AI-driven approach enhances the ability to predict and mitigate ozone pollution by considering broader meteorological patterns.