Abstract： To address the challenges of high pixel-level annotation costs, ambiguous object boundaries under limited labeling, easy breakage of slender structures, and loss of small targets in high-resolution remote sensing image semantic segmentation, we propose a semi-supervised semantic segmentation method that integrates visual distillation with structural awareness. Based on the FixMatch semi-supervised framework, we design a multi-teacher collaborative visual distillation architecture, leveraging pre-trained DINOv2 and CLIP as frozen teacher models to transfer general visual knowledge from the base models to the student network through feature alignment. We introduce a detail-preserving and semantic compression module that decouples highfrequency details from low-frequency structural features, thereby enhancing boundary perception and detection of small-scale objects. Additionally, we propose a dual-path parameterized context fusion module and an orientationaware strip attention module, which employ learnable channel-grouped parameters during decoding to preserve fine details and strengthen spatial continuity for linear features. Experimental results on the public LoveDA and DeepGlobe datasets demonstrate that with only 10% labeled data, our method achieves an average intersection over union (IoU) of 74.45% and an average F1 score of 84.86% on the DeepGlobe dataset, and an average IoU of 55.71% and an average F1 score of 69.43% on the LoveDA dataset—outperforming current mainstream semisupervised segmentation methods. Ablation studies validate the effectiveness of each module in improving boundary completeness and reducing structural fragmentation.