Prior Information-Guided Semi-Supervised Semantic Segmentation of Remote Sensing Images

Remote sensing semantic segmentation has broad applications in critical fields such as urban planning and environmental monitoring. Semi-supervised semantic segmentation methods tackle the time and cost challenges of annotating remote sensing images by using abundant unlabeled data to optimize the model. Many existing semi-supervised methods generate pseudo-labels by inputting unlabeled data into the network, and then use these labels to progressively improve the model’s segmentation performance. However, the imbalanced class distribution and complex spatial layout of remote sensing images may result in unreliable pseudo-labels generated by the networks. To address this issue, we propose a Prior Information Guided Network (PGNet) that integrates a Prior Feature Guided Module (PFGM) for reliable feature priors and a Prior Edge Extraction Module (PEEM) to provide edge prior information. By leveraging the lightweight Segment Anything Model (SAM), PGNet enhances the ability to extract valuable information from unlabeled data, enabling the generation of more reliable pseudo-label. The PFGM employs Symmetric Cross-Attention to more balancedly integrate the SAM prior features with the semantic features from the encoder. Additionally, the PEEM employs SAM to extract reliable edge priors from unlabeled data, guiding the feature maps during training for more accurate segmentation boundaries. We conducted extensive experiments on the ISPRS Vaihingen, ISPRS Potsdam, and LoveDA datasets. The results demonstrate that our PGNet outperforms existing state-of-the-art semi-supervised methods.