Spatially stratified modeling of intercity road freight flows in China

Understanding regional road freight flows is fundamental for logistics planning and collaborative regional development. However, comprehensive intercity road freight OD flow information remains incomplete in practice because conventional data sources are often aggregated or partially anonymized, while trajectory observations are fragmented by logistics organization, data standards, and privacy constraints. As a result, accurately and robustly inferring intercity road freight OD flows has become a critical task. Existing global modeling approaches usually rely on shared parameters and fail to capture the spatially stratified heterogeneity of road freight networks, where economic corridors, regional hubs, and peripheral areas exhibit distinct interaction patterns. This limitation leads to systematic inference bias, particularly under long-tailed flow distributions. To address this, we propose a Stratified Prediction (S2P) framework to improve the accuracy and robustness of intercity road freight OD flow inference by explicitly modeling spatially stratified heterogeneity. S2P first identifies latent strata through unsupervised clustering and multi-view feature analysis and then trains stratum-specific sub-models within a unified architecture with independent parameters. Each sub-model incorporates a feature weight learning module and a Huber-based piecewise loss function to capture heterogeneous interaction patterns. Applied to China’s intercity freight system, S2P reveals three functional subnetworks with distinct structural properties and distribution patterns. Compared with the best-performing baseline, S2P reduces RMSE and MAE by 10.3% and 20.6%, respectively, while also achieving the highest CPC and SCC values. It also maintains stable performance across different strata, demonstrating improved generalizability under heterogeneous and long-tailed road freight flows. These results demonstrate the value of explicitly incorporating stratified heterogeneity into intercity road freight OD flow inference and provide insights for differentiated logistics policies and sustainable freight planning. • Propose a method for inferring OD flows with spatially stratified heterogeneity. • Design stratum-specific model to capture distinct driving mechanisms and patterns. • Introduce adaptive feature weight learning to improve robustness across strata. • Enhance performance by addressing long-tail distribution and high-flow outliers. • Identify three freight subnetworks in China with tailored policy implications.