Estimating daily ground-level PM_2.5 in China with random-forest-based spatiotemporal kriging

Area of Science:

• Environmental Science
• Data Science
• Public Health

Background:

• Ambient fine particulate matter (PM2.5) is a significant contributor to cardiovascular and respiratory mortality.
• Traditional empirical models for PM2.5 estimation have limitations, including ignoring nonlinear relationships and assuming independent residuals.

Purpose of the Study:

• To develop and evaluate a hybrid approach for estimating daily PM2.5 concentrations.
• To improve the accuracy of PM2.5 prediction by integrating machine learning and geostatistical methods.

Main Methods:

• A hybrid model, Random Forest-based Spatiotemporal Kriging (RFSTK), was developed.
• The RFSTK model combines the strengths of the Random Forest (RF) model for capturing nonlinear interactions and spatiotemporal kriging for detailed dependence.

Main Results:

• The RFSTK model demonstrated superior performance compared to the original RF model.
• RFSTK achieved a 10-fold cross-validation R² of 0.881, MAE of 6.89 μg/m³, and RMSE of 11.48 μg/m³.
• In 2018, approximately 90.04% of China experienced daily PM2.5 exposure below the national standard of 75 μg/m³.

Conclusions:

• The proposed RFSTK model effectively estimates daily PM2.5 concentrations, outperforming traditional methods.
• The hybrid approach offers a generalizable solution for large-scale spatiotemporal mapping of ambient PM2.5.