Aiming at remote sensing image domain adaptation segmentation model challenges, including different feature distributions, different scales, and unequal distribution of categories, a remote sensing image domain adaptation segmentation model combining scale discriminator and attention was proposed. Firstly, for the different remote sensing image resolutions in different domains problem, in an adversarial-based discrimination network, in addition to the common feature discriminator, a scale discriminator was added to reduce the scale difference between the two domains. Secondly, the visual attention network (VAN) that considers spatial and channel attention was used as the feature extraction backbone network to improve the feature extraction capability. Finally, aiming at category samples' uneven distribution problem in remote sensing images, the FocalLoss loss function was introduced to improve the segmentation accuracy of categories with fewer samples by increasing attention to categories with fewer samples. Experimental results show that the proposed model can effectively alleviate the problems of remote sensing images with different features, scales, and unequal distribution categories. The proposed model significantly improved classification accuracy compared to the adversarial domain adaptation segmentation network based on a feature discriminator.
|