Machine learning models accurately predict ozone exposure during wildfire events.

Area of Science:

• Environmental science
• Epidemiology
• Machine learning

Background:

• Epidemiologists use prediction models to downscale air pollution exposure data where monitoring is scarce.
• Wildfires significantly impact air quality, necessitating accurate ozone exposure assessment.

Purpose of the Study:

• To compare the predictive accuracy of ten machine learning algorithms for ground-level ozone during wildfires.
• To evaluate model performance using a rigorous cross-validation technique that accounts for spatial and temporal data dependencies.

Main Methods:

• Ten machine learning algorithms were applied to predict daily 8-hour maximum average ozone concentrations.
• A leave-one-location-out cross-validation (LOLO CV) procedure was employed to assess model accuracy realistically.
• Performance was evaluated using root mean square error (RMSE) and R-squared (R²).

Main Results:

• Gradient boosting achieved the highest accuracy with the lowest LOLO CV RMSE (0.228) and highest R² (0.677).
• Random forest was the second-best performing model (LOLO CV R² of 0.661).
• LOLO CV provided less optimistic accuracy estimates than 10-fold cross-validation, especially for flexible models.

Conclusions:

• Gradient boosting is a highly accurate model for downscaling ozone exposure during wildfire events.
• Leave-one-location-out cross-validation offers more reliable accuracy estimates than k-fold cross-validation for spatiotemporal data.
• These models are suitable for interpolation but not for inferring causal effects or extrapolation to different conditions.