CollMatch: semi-supervised remote sensing image semantic segmentation via collaborative perturbation

In semi-supervised semantic segmentation of remote sensing images, consistency regularization (CR) improves model generalization by enforcing prediction consistency across different perturbed versions of unlabeled data. However, existing CR-based methods suffer from two major limitations: class confusion due to color-texture similarities, and the insufficient exploration of unlabeled data. To address these issues, we propose a multi-branch collaborative perturbation framework with two key components: region-aware context enhancement (RACE) and depth-adaptive feature consistency (DAFC). RACE employs a multi-scale regional masking strategy, combining grid and speckle masks to effectively occlude objects of different scales. This design utilizes mask consistency to compel the model to infer semantics from complementary contextual cues, thereby enhancing the model’s discriminative capability. DAFC introduces a hierarchical perturbation mechanism where intensity increases with feature depth, modulated by a dedicated function. By applying depth-adaptive consistency constraints at the feature level, it drives richer exploration of the latent feature space and improves the utilization of unlabeled data. Extensive experiments on four benchmark datasets show that our method achieves state-of-the-art mIoU, with gains over the baseline exceeding 2% on the MER, DFC22, and iSAID datasets.