Abstract: As an important research direction in the field of remote sensing technology, semantic segmentation of remote sensing images aims to segment remote sensing images at the pixel level and accurately classify them into predefined ground object categories. Traditional methods rely on a large number of pixel-level annotation data, but manual annotation is time-consuming and laborious. In order to solve this problem, a semi-supervised learning method is introduced. By combining a small number of labeled samples and a large number of unlabeled data for model training, the labeling requirements are significantly reduced. By analyzing and summarizing the related research on semi-supervised semantic segmentation of remote sensing images in recent years, the classification system of existing methods is systematically sorted out, and its advantages and disadvantages are deeply analyzed. Starting from the core ideas and technical strategies, the existing semi-supervised semantic segmentation methods of remote sensing images are systematically summarized and classified, and their innovations and limitations are discussed respectively. The datasets widely used in the research of semi-supervised semantic segmentation of remote sensing images are introduced. Based on the commonly used experimental settings and evaluation indicators, multi-method comparative analysis is carried out on different datasets. Finally, the future research trends of semi-supervised semantic segmentation of remote sensing images are discussed.

Key words: remote sensing image, semantic segmentation, deep learning, semi-supervised learning

摘要： 遥感图像语义分割作为遥感技术领域的重要研究方向，旨在对遥感图像进行像素级的分割并精确分类至预定义的地物类别中。传统方法依赖于大量的像素级标注数据，但人工标注耗时且费力。为解决这一问题，半监督学习方法被引入，通过结合少量标注样本和大量无标注数据进行模型训练，显著降低了标注需求。通过分析和总结近年来关于遥感图像半监督语义分割的相关研究，系统梳理现有方法的分类体系，并深入剖析其优势与不足：从核心思想和技术策略出发，对现有的遥感图像半监督语义分割方法进行了体系归纳和类别划分，并分别讨论了其创新与局限性；介绍了遥感图像半监督语义分割研究广泛使用的数据集；基于常用实验设置和评价指标，在不同数据集上开展了多方法对比分析；探讨了遥感图像半监督语义分割的未来研究趋势。

关键词: 遥感图像, 语义分割, 深度学习, 半监督学习