Cross-domain traffic forecasting through multimodal fusion and modality-aware learning

The continuous evolution of intelligent transportation systems (ITS) has driven a paradigm shift in traffic forecasting- from classical statistical models to advanced multimodal and cross-domain deep learning architectures. This paper presents a multimodal fusion and modality-aware learning framework evaluated on two real-world heterogeneous datasets: NGSIM (United States) and TiHAN-V2X (India). The framework integrates kinematic, contextual, and interaction modalities through early, intermediate, and late fusion strategies. For vehicle-type classification, early fusion achieves 0.9848 ± 0.0028 accuracy on TiHAN-V2X and 0.9996 ± 0.0001 on NGSIM(mean ± SD, n = 5 seeds). For short-term speed forecasting, TCN + Attention Early Fusion achieves a mean MAE of 6.638 ± 0.016 km/h at the 100 ms horizon and 10.293 ± 0.033 km/h at the 500 ms horizon across five random seeds, outperforming BiLSTM (6.680 ± 0.025 km/h) and TCN + Attention Late Fusion (6.701 ± 0.062 km/h) at all prediction horizons. A cross-dataset generalisation experiment applies NGSIM-trained classification models directly to TiHAN-V2X without retraining, producing physically consistent kinematic profile mappings across all seven TiHAN scenarios. The framework demonstrates strong missing-modality robustness, maintaining stable forecasting performance even when acceleration and lateral position are structurally absent. Late fusion consistently outperforms early fusion under modality dropout, while early fusion achieves the lowest forecasting MAE when modalities are well-aligned, providing complementary design guidelines for real-world ITS deployment. The results ( https://github.com/Jyoti-Yadav-PhD/multimodal-traffic-fusion ) establish that multimodal deep learning with spatial-temporal graph reasoning and attention-based temporal modelling is a viable foundation for next-generation, domain-generalizable traffic forecasting systems.