Urban ozone variability using automated machine learning: inference from different feature importance schemes

Area of Science:

• Atmospheric Chemistry
• Environmental Science
• Data Science

Background:

• Ground-level ozone is a harmful air pollutant and greenhouse gas, with urban emissions exacerbating pollution.
• Limited observational data and uncertain driving factors hinder ozone variability understanding in developing regions.

Purpose of the Study:

• To simulate ozone variability using machine learning (ML) and identify key influencing factors in a major Indian urban environment.
• To assess the effectiveness of Automated ML (AutoML) in modeling daily ozone concentrations.

Main Methods:

• Utilized ML models incorporating ozone precursors (NO2, NO, CO, C5H8, CH2O) from Copernicus Atmosphere Monitoring Service (CAMS) and meteorological data from ERA5.
• Employed Automated ML (AutoML) to optimize a deep learning model for daily ozone simulation.
• Applied feature importance schemes (SAGE, permutation importance) to understand model drivers.

Main Results:

• The AutoML-fitted deep learning model simulated daily ozone with a Root Mean Square Error (RMSE) of ~2 ppbv, capturing 84-88% of variability.
• Model performance was comparable to Random Forest (RF) and XGBoost models.
• Urban ozone simulation achieved RMSE of 2.5 ppbv and R² of 0.78 using the top four features identified by different importance schemes.

Conclusions:

• ML, particularly AutoML, offers a powerful tool for simulating and understanding urban ozone variability, complementing traditional models.
• Science-informed analysis of feature importance from multiple schemes is crucial for inferring variable roles in ozone photochemistry.
• This approach can enhance the accuracy of urban ozone simulation and forecasting, crucial for air quality management.