Objective: Change detection is highly sensitive to differential information in target features across multitemporal images and has been widely applied in various remote-sensing domains. It plays a crucial role in disaster emergency response and post-disaster damage assessment. However, most mainstream change detection methods are primarily designed for buildings, with limited applicability to roads. These methods struggle to capture the fine-grained and continuous change characteristics of roads. Furthermore, existing road detection approaches are predominantly single-temporal. Such approaches cannot perform end-to-end reconstruction of completely damaged or destroyed roads, which limits their performance in road damage detection. Moreover, publicly available datasets rarely include annotations for road damage information. The lack of specialized benchmark datasets for damaged roads therefore remains a challenge. To address these issues, this study developed a U-shaped bi-temporal transformer (U-BiFormer) for road damage change detection and constructed a bi-temporal damaged road (BTDR) dataset as a benchmark to validate the effectiveness of the proposed method. Methods: The proposed U-BiFormer consisted of three core modules: multiscale learning (MSL), residual transformer shortcut (RTS), and multilevel dense iteration decoding (MDID). MSL used the first three layers of ResNet-18 and a standard convolutional block to generate feature maps at four different scales. These multiscale feature maps enhanced and optimized the bi-temporal features, which were then fed into the RTS module. RTS consisted of two submodules: the Trans. block and the Res. block. The Trans. block modeled spatiotemporal global information within the bi-temporal features at different stages, containing high-level semantic information. The Res. block extracted difference maps that included low-level geographical localization information. By processing information at different hierarchical levels, RTS captured contextual dependencies at both long and short ranges, thereby improving performance. Instead of standard upsampling, MDID used dense blocks, which were particularly suitable for preserving road continuity. This module integrated feature maps from all preceding stages, enabling the fusion of coarse-grained and fine-grained features to support subsequent predictions. Finally, the proposed model employed separate prediction heads for the damage classification task and the change detection task. For damage classification, the class embedding vector was concatenated with semantic tokens to form new tokens, and the class embedding was then separated to predict the road damage level. For change detection, the reconstructed features were upsampled through several convolutional layers to generate the final visualization. Regarding the loss function, a combination of focal loss and Dice loss was used to achieve an optimal balance. Results: Experiments were conducted on the proposed BTDR dataset, including single-temporal road extraction experiments and bi-temporal change detection experiments. For the bi-temporal tasks, the results demonstrated that the proposed method performed well in damaged-category detection, with all evaluation metrics exceeding 82.00%. Furthermore, the method significantly outperformed mainstream models in the change detection task, achieving a mean intersection over union of 82.30%. In the single-temporal road detection task, the model also surpassed several mainstream road detection methods. Qualitative results visually confirmed the superiority of U-BiFormer. Conclusions: By constructing the specialized BTDR dataset and developing the U-BiFormer, this study achieves outstanding performance in road damage change detection and validates the effectiveness of each module. The proposed approach provides an efficient solution for rapid post-disaster response and reconstruction and offers significant engineering value.