An Ensemble Machine-Learning Model To Predict Historical PM_2.5 Concentrations in China from Satellite Data

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Satellite aerosol data provides historical PM2.5 (particulate matter with an aerodynamic diameter of 2.5 micrometers) estimates where ground monitoring is limited.
• Previous models often show reduced accuracy for predictions outside their training periods.

Purpose of the Study:

• To develop a reliable PM2.5 hindcast modeling system using an ensemble machine learning approach.
• To improve the accuracy of historical PM2.5 level assessments, particularly for periods with limited monitoring data.

Main Methods:

• Utilized multiple imputation for missing satellite data.
• Employed spatial clustering to divide China into seven regions, training separate machine learning models (Random Forest, GAM, XGBoost) for each.
• Developed a generalized additive ensemble model to integrate predictions from individual algorithms.

Main Results:

• The ensemble model accurately captured the spatiotemporal distribution of daily PM2.5, achieving a cross-validation R² of 0.79 and RMSE of 21 μg/m³.
• Subregion models outperformed national models, improving CV R² by approximately 0.05.
• Achieved improved out-of-range prediction accuracy compared to previous studies: R² = 0.58 (daily) and R² = 0.76 (monthly).

Conclusions:

• The proposed ensemble machine learning approach offers reliable PM2.5 hindcast capabilities.
• This system enables the construction of unbiased historical PM2.5 levels, crucial for environmental and health assessments.