Obtaining vertical distribution of PM_2.5 from CALIOP data and machine learning algorithms.

Area of Science:

• Environmental Science
• Atmospheric Science
• Remote Sensing

Background:

• Aerosol optical depth (AOD) is commonly used to estimate near-surface PM2.5 concentrations.
• However, traditional methods using total-column AOD cannot determine vertical PM2.5 distribution.
• Distinguishing AOD contributions at different altitudes is crucial for accurate air quality assessment.

Purpose of the Study:

• To compare machine learning models for estimating vertical PM2.5 concentrations from AOD.
• To analyze the influence of AOD at various altitudes on PM2.5 estimation.
• To derive and analyze the vertical distribution of PM2.5 concentrations.

Main Methods:

• Compared Extra Trees (ET), Random Forest (RF), Deep Neural Network (DNN), and Gradient Boosting Regression Tree (GBRT) models.
• Utilized altitude-resolved AOD data and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) data.
• Validated model performance using cross-validation and analyzed feature importance.

Main Results:

• The Extra Trees (ET) model demonstrated superior performance, achieving an R² of 0.85 and RMSE of 17.77 μg/m³.
• Bottom-layer AOD was found to be more important than upper-layer or total-column AOD for PM2.5 estimation.
• High PM2.5 concentrations were predominantly observed near the ground, decreasing with altitude, with significant downward trends in major Chinese regions.

Conclusions:

• The ET model effectively estimates vertical PM2.5 concentrations, outperforming other tested models.
• Vertical AOD profiles provide critical information for understanding PM2.5 distribution.
• The derived vertical PM2.5 data offers valuable insights for air pollution research and policy.