Abstract: In recent years, semantic segmentation of remote sensing images has been widely used. Although the method based on deep learning has greatly improved the semantic segmentation accuracy of remote sensing image, due to the diversity of remote sensing images（different geographic locations, terrain and weather conditions） and lack of pixel-level labels, the algorithm is difficult to apply across domains task. Retraining a new domain needs to collect the corresponding pixel-level labels, which consumes lots of human resources. So based on adversarial learning, a weakly-supervised domain adaptative algorithm with known image-level labels of the target domain is proposed：Firstly, using image-level labels, a multi-class domain discriminator is proposed to align each category of the target domain to the source domain adaptively. Secondly, to maintain the stability of the training process, a method to generate pixel-level pseudo-labels of the target domain based on the entropy value is proposed to supervise the feature extractor and improve the performance of model. Numerous experiments on ISPRS Vaihingen and ISPRS Potsdam datasets show that the improved algorithm is superior to other adaptive algorithms based on the adversarial learning domain in the semantic segmentation of remote sensing images.

Key words: semantic segmentation, remote sensing image, classification of ground objects, weakly-supervised domain adaptation

摘要： 近年来，遥感图像的语义分割得到广泛应用。虽然基于深度学习的方法极大程度地提高了遥感图像的语义分割精度，但由于遥感图像的多样性（不同的地理位置、地形和天气条件）以及像素级标签的缺乏，该算法难以适用于跨多个域的任务。而若重新训练新的域则需要消耗大量人力资源来收集相应的像素级标签。为了解决这一跨域问题，基于对抗学习提出了一种目标域图像级标签已知的弱监督域自适应方法：利用图像级标签，提出多类域判别器，使目标域各个类别自适应地对齐到源域；为保持训练过程的平稳，一种基于熵值产生目标域像素级伪标签的方法被提出以监督特征提取器，提升模型表现。在ISPRS Vaihingen与ISPRS Potsdam数据集上的大量实验表明，改进后的算法优于其他遥感图像语义分割中的基于对抗学习域自适应算法。

关键词: 语义分割, 遥感图像, 地物分类, 弱监督域自适应