High-Resolution Estimation of Daily PM_2.5 Levels in the Contiguous US Using Bi-LSTM with Attention.

Area of Science:

• Environmental Science
• Data Science
• Public Health

Background:

• Accurate estimation of surface-level PM2.5 concentrations is vital for public health.
• Existing models lack precision, particularly during high-concentration events.
• Limited open-source tools hinder broader application of PM2.5 estimation models.

Purpose of the Study:

• To develop an advanced deep learning model for improved PM2.5 concentration estimation.
• To enhance the precision of PM2.5 estimates, especially on high-concentration days.
• To provide an open-source framework and dataset for reproducible air quality research.

Main Methods:

• Developed a Long Short-Term Memory (LSTM) network with Attention mechanism.
• Integrated multiple data sources: in situ measurements, satellite data, and wildfire smoke density.
• Utilized temporal dynamics of air quality for improved estimation.

Main Results:

• Achieved a 2.2% improvement in overall Root Mean Square Error (RMSE).
• Demonstrated a 9.8% reduction in RMSE on high-concentration days.
• Created a comprehensive PM2.5 dataset for the contiguous US (2005-2021).

Conclusions:

• The novel deep learning model significantly improves PM2.5 estimation accuracy.
• The model excels in predicting high PM2.5 concentration events.
• The release of an open-source framework promotes reproducibility and further research in air quality studies.