CPDM: Class Prototypes With Dynamic Multidomain Mixed Perturbation Consistency for Semi-Supervised Image Segmentation

To address the technical bottleneck in remote sensing (RS) image semantic segmentation caused by the high annotation cost and reliance on large amounts of pixel-level labels, this study proposes a novel semi-supervised RS image semantic segmentation method that alleviates the limitations of existing approaches in limited perturbation space, single consistency learning, and noisy pseudo-labels. Specifically, for unlabeled RS images, a multi-domain CutMix hybrid strong augmentation strategy (MCS) is proposed. This strategy performs adaptive region masking operations, based on the traditional CutMix strategy, to enlarge the model’s perturbation space. In addition, considering that single weak-to-strong consistency learning is not enough to help the model capture richer contextual information, this study introduces prototype-based prediction result consistency (PBC) to reduce the model’s dependence on weak-to-strong consistency. Finally, a class-adaptive entropy mapping (CAEM) method is proposed to refine noise filtering for each type of result and guide the model to optimize in the right direction. Extensive experimental results show that the proposed method significantly outperforms existing advanced methods on six mainstream RS image datasets: Vaihingen, MER, MSL, GID-15, iSAID, and DFC22. The code is available at https://github.com/ZhuangQingYuan/CPDM.git.