A two-stage domain adaptive remote sensing image semantic segmentation network combined with self-training

Unsupervised domain adaptive (UDA) is a popular technology to solve the semantic segmentation of remote sensing images. UDA can alleviate the need for large-scale pixel-level labeling by pre-training on the labeled source domain and verifying on the target domain. However, domain offset exists between different data sets, which will greatly affect the final performance of the semantic segmentation model. Based on this, this paper proposes a segmented interdomain adaptive network framework of dual-path image transformation, which combines self-training with interdomain adaptive segmentation. Firstly, the image style conversion is used to reduce the domain offset of the image level, and the dual-path method is used to promote each other to improve the image conversion quality and the pseudo-label quality of the target domain. Secondly, the pseudo-tags of the first-stage training are used to prioritize, and the target domain is divided into easy and difficult to partition for selective retraining. By adversarial learning, the difficult-to-segment features are aligned with easily-segmented features, and the influence of domain migration is gradually reduced by adaptive repetition in the domain, thus improving the segmentation performance of the network.