To mitigate the problem of diminished model representation capabilities stemming from the limited availability of labeled data in runway detection for polarimetric synthetic aperture radar (PolSAR) images, a bidirectional, collaborative, semi-supervised teacher-student model is introduced. Specifically, an assistant teacher module has been designed to facilitate joint model training through the construction of distillation and feedback losses, thus overcoming the hierarchical constraints of conventional unidirectional distillation. Feedback on the models' inference results is provided by the assistant teacher module to extract previously unexploited feature information. Furthermore, directional feature vectors are generated using peer-level feature maps, and a directional loss is constructed to aid in the efficient training of the student model. Experiments conducted on the U.S. UAVSAR dataset demonstrate that, under conditions of limited labeled data, the proposed method achieves a runway area detection accuracy of 83.11%, which represents improvements of 15.63%, 6.46%, and 17.25% compared to the Unet, D-Unet, and Unet++ models, respectively.