<p>The damage identification of arch bridges plays a vital role in the structural condition assessment and maintenance decision-making. However, traditional deep learning-based methods confront the issue of insufficient accuracy and high requirements of training samples. To this end, this paper proposes an arch bridge damage identification method based on MobileViT and transfer learning. First, a novel damage index (i.e., cross-correlation matrix) is proposed to process the original acceleration response of arch bridges, by which the damage dataset is constructed to characterize more structural state information. Second, a lightweight MobileViT model combined Transformer and lightweight convolutional neural network is built to mitigate the computational burden and improve the recognition ability. Subsequently, the transfer learning strategy is employed to train the MobileViT model, by which the sample requirement is significantly reduced. Then, the trained model is used for the damage identification of arch bridges. Finally, the feasibility and superiority of the proposed method are verified numerically and experimentally. The influence of several key factors (e.g. sample size, sensor failure rate) on damage identification is analyzed. The results demonstrate that the proposed method achieves the highest identification accuracy with the lightest model structure compared with traditional methods (i.e., AlexNet, ResNet50, InceptionV3 and VGG16). The proposed method obtains an accuracy of 95.9% in the experiment, and achieves an improvement of 6.2% compared with VGG16. In addition, the proposed method shows excellent robustness under the influence of factors like sensor failure and noise disturbance.</p>