<p>MobileViT v2, a Transformer-CNN hybrid deep learning model, was restructured and successfully applied to predict urban pluvial flood maximum water depth for the first time. The hybrid model leverages the advantages of both Transformers and CNNs to enhance global and local modeling capabilities, thereby achieving high prediction accuracy even on unseen terrains. Furthermore, we proposed a data augmentation approach to satisfy the substantial data requirements of the Transformer blocks and mitigate overfitting caused by limited training data. Moreover, we utilized an attribution method to analyze the effectiveness and physical plausibility of both the hybrid model and the data augmentation, thereby enhancing the interpretability of the proposed framework. Results indicate that the hybrid model with data augmentation demonstrated a significant improvement in performance compared to the standard ViT and a modest enhancement compared to representative CNN baselines. These findings demonstrate the hybrid model’s superior generalization capability and greater potential for further improvements. Additionally, the results validate data augmentation as an easy-to-implement and widely applicable approach for enhancing prediction accuracy with limited data in flood prediction tasks. Finally, this study presents a promising method for enhancing prediction accuracy in flood prediction tasks and represents the successful application of hybrid models to this task across diverse unseen spatial terrain patches.</p>