Robust prediction of hourly PM_2.5 from meteorological data using LightGBM

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Historical fine particulate matter (PM2.5) data is vital for assessing environmental, health, and climate impacts.
• Existing satellite-based PM2.5 estimations suffer from missing data, low frequency, and poor predictive power.

Purpose of the Study:

• To develop a robust model for accurate, high-resolution historical PM2.5 data reconstruction.
• To overcome limitations of current satellite-based PM2.5 estimation methods.

Main Methods:

• Employed a novel feature engineering approach incorporating spatial effects from meteorological data.
• Developed and validated a LightGBM model for PM2.5 prediction.
• Utilized spatial features to construct hourly gridded PM2.5 networks.

Main Results:

• Achieved unprecedented predictive capacity for PM2.5 on hourly (R²=0.75), daily (R²=0.84), monthly (R²=0.88), and annual (R²=0.87) timescales.
• Demonstrated the model's ability to create hourly gridded PM2.5 networks.
• Highlighted potential for enhanced accuracy with regional meteorological data integration.

Conclusions:

• The developed LightGBM model shows significant potential for reconstructing historical PM2.5 datasets and real-time gridded networks.
• The model's high spatial-temporal resolution capabilities are valuable for environmental and climate change research.
• Generated datasets can be assimilated into models for long-term re-analysis, including aerosol-physical process interactions.