A multi-resolution ensemble model of three decision-tree-based algorithms to predict daily NO₂ concentration in...

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Public Health

Background:

• Accurate Nitrogen Dioxide (NO2) exposure data is crucial for understanding its health impacts.
• Existing exposure maps often lack the necessary spatiotemporal resolution.
• High-resolution NO2 mapping is needed for effective environmental and health management.

Purpose of the Study:

• To develop and validate a multi-stage, multi-resolution ensemble model for predicting daily NO2 concentrations in France.
• To achieve high spatiotemporal resolution (200m in urban areas) for NO2 exposure assessment.
• To provide reliable NO2 exposure estimates for epidemiological studies.

Main Methods:

• A three-stage ensemble modeling approach was employed, integrating satellite data, land cover, and traffic information.
• Stage 1: Predicted NO2 total column density from satellite observations.
• Stage 2 & 3: Used generalized additive models with decision-tree algorithms (Random Forest, XGBoost, CatBoost) to predict NO2 concentrations at 1km and 200m resolutions, respectively, incorporating spatio-temporal blocking for robust validation.

Main Results:

• The 1km resolution model achieved a cross-validated R² of 0.83, while the 200m urban model reached an R² of 0.69.
• The ensemble approach demonstrated good predictive performance and minimized errors in daily NO2 concentration predictions.
• The model successfully captured NO2 variations across continental France from 2005 to 2022.

Conclusions:

• The developed multi-stage ensemble model provides unprecedented high-resolution daily NO2 exposure maps for France.
• This approach ensures robust performance estimation and accurate predictions, even when individual algorithms falter.
• The generated exposure estimates are valuable resources for future research on NO2-related health effects.