A Spatiotemporal Decomposition Framework for Temporal Persistence and Regional Transport in PM_2.5 Variability

Area of Science:

• Environmental Science
• Data Science
• Air Quality Modeling

Background:

• Fine particulate matter (PM2.5) significantly impacts environmental health in the U.S.
• Existing models struggle to differentiate local PM2.5 persistence from regional transport effects.
• Accurate spatiotemporal prediction of PM2.5 is essential for public health and environmental policy.

Purpose of the Study:

• To develop an advanced spatiotemporal model for PM2.5 prediction.
• To explicitly separate temporal dynamics from spatial predictors in PM2.5 modeling.
• To improve the characterization of short-term PM2.5 variability.

Main Methods:

• Integration of a bidirectional long short-term memory (BiLSTM) network for temporal dependencies.
• Utilization of Kolmogorov-Arnold Networks (KAN) with splines for nonlinear spatial gradients.
• Application of a case study in Texas to evaluate model performance.

Main Results:

• Significant reductions in RMSE (39.6%) and MAE (41.9%) compared to conventional LSTM.
• A notable increase in R-squared (16.7%) indicating improved predictive power.
• High accuracy (precision 0.93, recall 0.96) in county-scale next-day exceedance assessment.

Conclusions:

• The proposed model effectively separates temporal persistence from spatial transport influences for PM2.5.
• This approach offers superior characterization of short-term PM2.5 variations.
• The framework is adaptable for regions with varying air quality monitoring coverage.