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

Fine particulate matter (PM_2.5) remains a major contributor to environmental health burden in the United States. Its concentrations vary across space and time due to the combined influence of local temporal persistence and regional transport processes. Many existing modeling approaches represent spatial and temporal dependencies within a single predictive structure, limiting the ability to distinguish site-specific temporal persistence from regional transport and emission-related influences. In this study, we develop a spatiotemporal modeling strategy that explicitly separates temporal dynamics from spatial predictors by integrating a bidirectional long short-term memory (BiLSTM) network with Kolmogorov-Arnold Networks (KAN). The BiLSTM component captures forward and backward temporal dependencies in site-level PM_2.5 time series, while two spline-based KAN layers represent nonlinear spatial gradients using distance-weighted PM_2.5 and ancillary predictors. This structure enables independent modeling of temporal persistence and spatial transport processes while retaining flexibility in nonlinear feature representation. Using Texas as a case study, the proposed model reduced RMSE by 39.6% and MAE by 41.9%, and increased R² by 16.7% relative to a conventional LSTM, while more accurately capturing high-concentration episodes. Site-level predictions were further aggregated to support county-scale next-day exceedance assessment, achieving precision of 0.93 and recall of 0.96 across six major counties. By separating temporal persistence from spatial transport influences, this approach provides improved characterization of short-term PM_2.5 variability and provides a modeling framework applicable to regions with heterogeneous monitoring coverage.