Multimodal Spatio-Temporal Heterogeneous Graph Transformer Network for citywide zone-level crash count prediction from multimodal urban data

This study proposes MSTHGTNet, a Multimodal Spatio-Temporal Heterogeneous Graph Transformer Network for zone-level crash count prediction. Zone-level crash risk is driven by dynamic exposure, spatial spillovers, and road-network propagation, under strong spatiotemporal and cross-modal data heterogeneity. MSTHGTNet addresses these challenges with three coordinated modules: a Transformer temporal encoder that models long-range dependencies in multimodal sequences (taxi-derived mobility, weather, and local/neighbor crash histories), an HGT-based road-network encoder that preserves road-type semantics for risk propagation, and a cross-modal attention fusion mechanism that conditions temporal exposure representations on spatial–static context by attending to heterogeneous road and zone attributes. Using 2022 New York City multimodal data, MSTHGTNet achieves MAE = 0.1267 and RMSE = 0.2902, reducing MAE by 17.99% relative to GMAN, the strongest non-proposed baseline, and outperforming both classical and recent spatiotemporal graph/attention-based models. Ablation and module-combination studies show that relation-aware heterogeneous road encoding and attention-based cross-modal fusion consistently improve over homogeneous-graph and simple-fusion variants, and that temporal and spatiotemporal traffic-demand features are the most influential modalities. Cross-city evaluations further report comparable performance in Chicago and San Francisco, with each city favoring a weekly aggregation window.