Remote sensing image semantic segmentation is crucial for various applications, but its effectiveness is hindered by the labor-intensive and time-consuming process of pixel-level annotation. Semi-supervised learning (SSL) offers a promising solution by leveraging abundant unlabeled data with limited labeled samples. However, a significant challenge in existing SSL methods lies in generating reliable pseudo-labels, particularly around complex boundaries, which often suffer from uncertainty. To address this, we propose a novel semi-supervised semantic segmentation network that enhances pseudo-label accuracy through the integration of more reliable edge information. Our contributions include: 1) an innovative edge detection branch that combines Dice Loss and BCE Loss to address class imbalance and boundary blurring, enabling the model to learn fine-grained and accurate edges simultaneously; and 2) a dynamic Exponential Moving Average (EMA) update strategy for the teacher model. This strategy adaptively adjusts the teacher model’s parameter update rate during training, thereby significantly accelerating and optimizing the student model’s learning process. Extensive experiments conducted on the LoveDA and ISPRS Vaihingen datasets demonstrate that our proposed method significantly outperforms existing state-of-the-art approaches under limited labeled data conditions, achieving superior segmentation accuracy and consistency.