An Alternating Guidance With Cross-View Teacher-Student Framework for Remote Sensing Semi-Supervised Semantic Segmentation

The semantic segmentation of remote sensing images is crucial for Earth observation. The semi-supervised semantic segmentation method can effectively reduce the dependence of the training process on labeled data. Among them, the semi-supervised semantic segmentation method based on the teacher-student paradigm is currently one of the most mainstream methods. However, the issue of weight coupling has constrained further performance improvements. This article proposes an alternating guidance method that combines cross-view learning to improve the teacher-student paradigm and enhance the semantic segmentation performance of remote sensing images. The student model is designed by using two decoders with the same architecture but independently updated parameters. Two decoders process the input obtained after image and feature level perturbations. This allows the student model to generate unique feature representations and enhances its learning capability. The teacher model uses two decoders to construct an alternating supervision mechanism. The two decoders of the teacher model take turns outputting pseudo-labels to guide the training process of the student model. This alternating supervision strategy can provide richer supervision signals for student model while helping to alleviate weight coupling between teacher and student. The experiments on two remote sensing image datasets show that compared with the state-of-the-art (SOTA) semi-supervised semantic segmentation methods, the method proposed demonstrates excellent competitiveness.