Travel mode identification: a spectral-attention enhanced approach with geographic context

Inferring travel mode distribution is key to understanding individual travel behaviour and reconstructing travel chains, which is crucial for personalized travel services. An end-to-end model architecture is proposed to enhance discriminative feature representation through context-aware geographic embeddings that integrate bus stop and intersection information. An innovative feature encoding framework augmented with spectral and attention mechanisms is developed: periodic movement patterns are extracted through spectral analysis, while spatiotemporal dependencies are jointly modeled via combined self-attention and positional attention. Furthermore, a hybrid Transformer-CRF decoding layer is constructed to ensure precise point-wise classification, with plausible mode transitions being modeled through state transition probabilities, significantly enhancing trajectory coherence. Experiments on real GPS data show our model achieves an 88.9% accuracy, outperforming existing methods. Ablation and comparison tests validate the effectiveness of our approach, confirming its superiority in travel mode identification.