Related Experiment Video
Updated: Feb 20, 2026

Split Point Analysis and Uncertainty Quantification of Thermal-Optical Organic/Elemental Carbon Measurements
Published on: September 7, 2019
Enhancing Estimation of Fine Particulate Matter Chemical Composition across North America by Including Geophysical A
Siyuan Shen1, Aaron van Donkelaar1, Nathan Jacobs2
1Department of Energy, Environmental, and Chemical Engineering, Washington University in St. Louis, St. Louis, Missouri 63130, United States.
This study enhances estimates of fine particulate matter (PM2.5) and its components across North America using convolutional neural networks (CNNs). New validation methods reveal model performance and uncertainty, improving air quality research.
Area of Science:
- Environmental Science
- Atmospheric Chemistry
- Data Science
Background:
- Exposure to fine particulate matter (PM2.5) is a major global health risk.
- Accurate characterization of PM2.5 chemical composition is crucial for health studies and environmental management.
- Existing methods for estimating PM2.5 concentrations have limitations in spatial coverage and accuracy.
Purpose of the Study:
- To improve estimates of total PM2.5 mass concentration and its chemical components across North America.
- To develop and apply advanced machine learning models (CNNs) for enhanced PM2.5 estimation.
- To introduce and utilize a novel cross-validation technique (BLISCO) for robust model evaluation, especially in remote areas.
Main Methods:
- Development and optimization of convolutional neural networks (CNNs) integrating satellite, simulation, and ground monitor data.
- Application of CNNs to estimate monthly PM2.5 and component concentrations across North America (2000-2023).
- Implementation of traditional 10-fold spatial cross-validation and a new Buffered Leave Isolated Sites and Clusters Out (BLISCO) method for model validation and uncertainty assessment.
Main Results:
- CNNs demonstrated significant agreement with traditional cross-validation for total PM2.5 and major components (e.g., R² for total PM2.5 = 0.82, sulfate = 0.98).
- BLISCO cross-validation highlighted potential overestimation of performance and underrepresentation of uncertainty by traditional methods.
- Incorporating chemical transport model (GEOS-Chem) data improved CNN performance in BLISCO, notably for nitrate (R² 0.51 to 0.81) and ammonium (R² 0.27 to 0.67).
Conclusions:
- The developed CNN approach provides improved estimates of PM2.5 and its chemical components across North America.
- The BLISCO validation method offers a more realistic assessment of model performance and uncertainty, particularly for extrapolation.
- Distance from monitoring sites is a key factor influencing the uncertainty of PM2.5 estimates in remote regions.
More Related Videos
Related Concept Videos
Estimation of the Physical Quantities
Application of Linearization and Approximation

