Stacking machine learning model for estimating hourly PM_2.5 in China based on Himawari 8 aerosol optical depth data

Area of Science:

• Environmental Science
• Atmospheric Science
• Remote Sensing

Background:

• Traditional methods for estimating surface particulate matter (PM2.5) concentrations rely on polar orbit satellite data and meteorological factors.
• Existing methods lack high temporal resolution due to limitations in satellite observation capabilities.

Purpose of the Study:

• To develop a high-temporal-resolution method for estimating PM2.5 concentrations using geostationary satellite data.
• To assess the spatial and temporal distribution of PM2.5 in Central and Eastern China.

Main Methods:

• Utilized 1-hour aerosol optical depth (AOD) data from the Himawari 8 geostationary satellite.
• Developed a stacking machine learning model integrating AdaBoost, XGBoost, and random forest, combined with multiple linear regression.
• Applied the model to estimate hourly PM2.5 concentrations across Central and Eastern China.

Main Results:

• The proposed stacking model demonstrated superior performance over single models, achieving an average R² of 0.85 and RMSE of 17.3 μg/m³.
• Peak model precision (R²=0.92, RMSE=12.9 μg/m³) was observed at 14:00 local time.
• The North China Plain was identified as the most polluted area, with an annual daytime average PM2.5 concentration of 58 μg/m³.
• PM2.5 pollution levels were highest in winter, averaging 73 μg/m³.

Conclusions:

• The stacking model effectively estimates hourly PM2.5 concentrations with high accuracy and temporal resolution.
• Geostationary satellite AOD data is a valuable resource for improving PM2.5 monitoring.
• Seasonal and regional variations in PM2.5 pollution in China were clearly delineated.