Improved retrieval of PM2.5 from satellite data products using non-linear methods

Area of Science:

• Environmental Science
• Atmospheric Science
• Remote Sensing

Background:

• Current air quality data relies heavily on surface measurements, limiting areal coverage.
• Satellite observations offer potential for broader spatial data on air quality.
• Particulate matter (PM) exposure assessment is crucial for public health.

Purpose of the Study:

• To compare statistical methods for retrieving daily PM2.5 concentrations using satellite data.
• To evaluate the effectiveness of Linear Regression (LR), Generalized Additive Models (GAM), and Multivariate Adaptive Regression Splines (MARS).
• To identify the optimal method for improving PM2.5 estimations in the San Joaquin Valley, California.

Main Methods:

• Utilized satellite products MODIS-AOD and OMI-AAI for PM2.5 retrieval.
• Compared three statistical models: Linear Regression (LR), Generalized Additive Models (GAM), and Multivariate Adaptive Regression Splines (MARS).
• Assessed model performance using correlation coefficients (R²).

Main Results:

• Simple Linear Regression (LR) showed poor performance (R² ≈ 0.2) in the western USA.
• Both GAM (R² = 0.61) and MARS (R² = 0.71) significantly outperformed LR.
• MARS demonstrated a slight performance advantage over GAM and superior computational efficiency.

Conclusions:

• Multivariate Adaptive Regression Splines (MARS) is a highly effective method for retrieving PM2.5 from satellite aerosol products.
• Improved PM2.5 retrievals can supplement sparse ground monitoring data.
• This approach supports air quality model evaluation and epidemiological exposure assessments.