关键词: 域自适应, 多源遥感, 图像分类, 合成孔径雷达, 标签噪声

Abstract: The Convolutional Neural Network (CNN) is currently the mainstream algorithm for optical image target recognition. However, due to the complexity of Synthetic Aperture Radar (SAR) imaging, annotated data are scarce and difficult to obtain, making it challenging to meet the training sample requirements of CNNs. Training models with simulated/optical images and then bridging the domain gap between simulated/optical images and real SAR images through Unsupervised Domain Adaptation (UDA) technology has emerged as a feasible solution. Nevertheless, most existing UDA methods assume that all source domain features are transferable, overlooking the fact that certain domain-specific features may cause negative transfer, while pseudo-label noise can also adversely affect domain alignment. To address the above issues, this paper proposes a domain adaptation framework that integrates feature decoupling and label noise suppression. The method employs a feature decoupling module to separate sample representations into domain-shared features and domain-specific features, aligning only the shared features to reduce the risk of negative transfer. Additionally, a weighted generalized cross-entropy loss function is introduced to mitigate the interference of label noise generated during pseudo-label iteration. Experimental results demonstrate that the proposed method effectively improves recognition accuracy and domain adaptation robustness in cross-domain SAR target recognition tasks.

Key words: domain adaptation, multi-source remote sensing, image classification, synthetic aperture radar, label noise