一种用于遥感影像语义分割对抗域适应方法

为利用已有标注的影像数据集实现对未标注遥感影像的语义分割,提出一种对抗域适应的方法。首先在生成对抗网络的基础上,利用基于香农熵的不确定图,进行域间的对抗学习,实现已标注的影像(源域)与未标注的影像(目标域)之间的迁移学习;其次为进一步提升模型的无监督学习能力,使用基于伪标签提纯的自学习策略。为验证所提方法的有效性,使用ISPRS提供的IRRG波段的Vaihingen数据集与RGB波段的Potsdam数据集进行实验。实验结果表明,与典型的域适应方法相比,该方法可以有效地提升网络的泛化能力,进而提高模型在目标域上的分割精度。

An Adversarial Domain Adaptation Method for Semantic Segmentation of Remote Sensing Imagery

To use the existing labeled remote sensing datasets for semantic segmentation of unlabeled ones,a domain adaptation method based on a generative adversarial network was proposed in the paper.Firstly,the uncertainty maps calculated through Shannon Entropy was used to domanial adversarial learning so as to realize the transfer learning between labeled images(source domain)and unlabeled images(target domain).Secondly,a self-learning strategy was utilized basing on rectifying pseudo label to further improve the unsupervised learning performance.In order to verify the effectiveness of the proposed method,experiments were performed using the Vaihingen datasets and the Potsdam datasets.The Vaihingen datasets composed of IRRG-band images and the Potsdam datasets composed of RGB-band ones were used which were provided by ISPRS.The experimental results demonstrate that the proposed method could efficiently improve the generalization ability of the network and the segmentation accuracy in the target domain,compared with other typical domain adaptation methods.