<p>Accurate travel mode identification from GPS trajectories is critical for urban planning, traffic management, and location‐based services. Traditional two‐stage approaches, which first segment trajectories and then classify segments, suffer from segmentation errors and system complexity. Existing one‐stage deep‐learning methods often neglect local spatiotemporal variations, multi‐scale contextual features, and cross‐user generalization. To address these challenges, we propose DIT, an end‐to‐end framework that integrates Dynamic Spatio-Temporal Convolution for adaptive local encoding, Inception modules for multi‐scale feature fusion, and a Transformer encoder for global context modeling. We further enrich point‐level representations with temporal and public‐transit POI features and incorporate a contrastive learning loss to enhance class separability. On the GeoLife benchmark, DIT achieves 0.8582 accuracy and 0.8576 F1-score, outperforming the previous one‐stage state‐of‐the‐art by 1.06%. To evaluate cross‐user robustness, we simulate source-target domain shifts; DIT maintains accuracy 15.57% above the baseline. Moreover, embedding a Domain Adversarial Neural Network module yields an additional accuracy gain of 8.02% and F1 improvement of 6.59% on the target domain. Sensitivity and ablation studies confirm the individual contributions of each module. These results demonstrate DIT’s potential for scalable, high‐precision travel mode recognition across diverse real‐world scenarios.</p>