Abstract: The shortage of systematic and complete highway sample datasets makes it difficult for the traditional deep learning paradigm to meet the application requirements.Therefore，to address this problem，this paper proposed a two-stage scene recognition framework under small-sample conditions，with easily accessible remote sensing data for training tests，and then finally applied to highway scenes. The small-sample learning started with obtaining a priori knowledge from a large-scale base class dataset，learning the base model，and then generalizing the model to new classes that did not appear in the training process or had few training samples. In the first stage of the framework，we introduced a multi- task model to learn the intrinsic features across semantic classes from two auxiliary tasks，and then in the second stage，we implemented joint prediction of labeled and unlabeled data based on label propagation. Extensive experiments showed that the scene recognition method proposed in this article achieved a classification accuracy of 80. 58%，which was an improvement of 13. 24% and 10. 69% compared to SIB and CAN+T，respectively. It performed excellently in the test dataset with a classification accuracy of 69. 37%. This method can be used for scene recognition and intelligent warning tasks for engineering vehicle driving safety in various highway construction projects.

Key words: few-shot scene classification, road construction, safety inspection, remote sensing image