Predicting monthly high-resolution PM_2.5 concentrations with random forest model in the North China Plain

Area of Science:

• Environmental Science
• Public Health
• Atmospheric Science
• Data Science

Background:

• Global public health is significantly impacted by fine particulate matter (PM2.5) exposure.
• Epidemiological studies in developing nations are hampered by a scarcity of PM2.5 monitoring data.
• Reliable historical PM2.5 exposure data, especially before 2013, is rare in China.

Purpose of the Study:

• To develop a high-performance machine learning model for estimating monthly PM2.5 levels in North China Plain.
• To generate reliable historical PM2.5 exposure data for epidemiological research.
• To assess PM2.5 concentrations in specific polluted regions of China.

Main Methods:

• Developed a random forest model using Multi-angle implementation of atmospheric correction (MAIAC) aerosol optical depth (AOD), meteorological data, land cover, and ground PM2.5 measurements (2013-2015).
• Applied a multiple imputation method to address missing AOD values.
• Evaluated model performance using 10-fold cross-validation and validated with a separate time period (2016).

Main Results:

• The model achieved high performance with a cross-validation R² of 0.88 and relative prediction error (RPE) of 18.7%.
• Validation demonstrated accurate prediction of historical PM2.5 concentrations at monthly (R²=0.74, RPE=27.6%), seasonal (R²=0.78, RPE=21.2%), and annual (R²=0.76, RPE=16.9%) levels.
• The annual mean predicted PM2.5 concentration from 2013-2016 was 67.7 μg/m³, with Southern Hebei, Western Shandong, and Northern Henan identified as the most polluted areas.

Conclusions:

• The developed machine learning model provides a computationally efficient and high-resolution method for estimating historical PM2.5 concentrations.
• This approach can generate reliable data crucial for epidemiological studies on PM2.5 health effects in China.
• The findings highlight significant PM2.5 pollution hotspots requiring targeted public health interventions.