RSProtoSemiSeg: Semi-supervised semantic segmentation of high spatial resolution remote sensing images with probabilistic distribution prototypes

Semi-supervised semantic segmentation of high spatial resolution remote sensing images aims to mitigate the reliance on labeled data by using limited labeled data alongside extensive unlabeled data. This approach significantly reduces the dependency on labeled images and annotation costs and mitigates the challenge of obtaining large-scale labeled datasets. Current semi-supervised methods for remote sensing images semantic segmentation mainly focus on improving pseudo-label quality. However, the inherent noise in pseudo-labels remains a critical issue, leading to persistent inaccuracies. Furthermore, the high intra-class variance in such image complicate pixel-wise label propagation, exacerbating pseudo-label errors and substantially constraining segmentation accuracy. To tackle these limitations, we propose a contrastive learning framework based on a mixture of Gaussian mixture distributions. Our approach uses a mixture probability distribution prototype predictor to adaptively regulate the influence of intra-class prototypes on pixel representations, mapping features to a multivariate Gaussian model to mitigate pseudo-label inaccuracies. We also introduce a novel mixture contrastive loss function to guide pseudo-labeled pixels towards intra-class prototypes and away from inter-class prototypes while repelling, thereby enhancing representation fidelity. Extensive experiments on three remote sensing semantic segmentation datasets demonstrate the efficacy of our approach. Compared to baseline models, our method achieves mIoU improvements ranging from 0.26% to 1.86%. Compared to the state-of-the-art DWL model, it achieves even higher mIoU improvements, ranging from 0.37% to 2.36%. These results confirm the effectiveness of our approach over existing semi-supervised segmentation methods. The code will be available at https://github.com/aitointerp/rsprotosemiseg.