Accurate runoff prediction is critical for effective water resource management, particularly in data-scarce watersheds. Traditional runoff prediction models, such as process-driven approaches, often require extensive hydrometeorological data, which are challenging to obtain in regions with limited data availability. To address this issue, we propose RSformer, an enhanced Transformer-based model for runoff prediction. RSformer improves the embedding module of the Transformer to enhance the semantic richness of input data and employs segmentation processing combined with self-attention mechanisms to capture local features while preserving variable independence. Furthermore, a convolutional neural network (CNN) module is integrated to extract spatial features, and an extreme value loss function is designed to improve the model’s performance in predicting extreme events. Experimental results demonstrate that RSformer outperforms the traditional Transformer model by 18% in runoff prediction tasks across five data-scarce watersheds. This study provides a robust solution for runoff prediction in data-scarce regions.