CLR-DLR: a semi-supervised framework for high-fidelity remote sensing segmentation

Remote sensing images are crucial for large-scale earth observation, providing vital data for land use, urban development, and environmental monitoring. However, the complexity of semantic segmentation tasks and the need for extensive pixel-level annotations pose significant challenges. To overcome these obstacles, researchers have explored semi-supervised approaches, consistency regularization, pseudo-labeling, and generative adversarial network (GAN)-based methods. Despite their potential, these methods often struggle with issues like over-smoothing, noisy label propagation, and training instability. To address this problem, this article introduces a novel semisupervised model that combines pseudo-labeling with consistency regularization to address multiscale challenges and domain gaps. By implementing contextual label realignment (CLR) in the label space, we refine noisy labels, while denoising latent restoration (DLR) in the feature space uses a conditional diffusion model to tackle latent uncertainty. Our experimental results show that this framework effectively manages diverse imaging conditions and blurred boundaries, achieving superior robustness and accuracy with limited annotations. In addition, a polygonization-based boundary refinement enhances segmentation quality for complex object shapes ensuring coherent delineation.