End-to-End Unsupervised Style and Resolution Transfer Adaptation Segmentation Model for Remote Sensing Images

For remote sensing image unsupervised domain adaptation, there are differences in resolution except for feature differences between source and target domains. An end-to-end unsupervised domain adaptation segmentation model for remote sensing images is proposed to reduce the image style and resolution differences between the source and target domains. First, a generative adversarial-based style transfer network with residual connection, scale consistency module, and perceptual loss with class balance weights is proposed. It reduces the image style and resolution differences between the two domains and maintains the original structural information while transferring. Second, the visual attention network (VAN) that considers both spatial and channel attention is used as the feature extraction backbone network to improve the feature extraction capability. Finally, the style transfer and segmentation tasks are unified in an end-to-end network. Experimental results show that the proposed model effectively alleviates the performance degradation caused by different features and resolutions. The segmentation performance is significantly improved compared to advanced domain adaptation segmentation methods.