Semisupervised Cross-Domain Remote Sensing Scene Classification via Category-Level Feature Alignment Network

In recent years, remote sensing scene classification has obtained much attention owing to its widespread applications. Nevertheless, the existing deep learning-based scene classification methods suffer from domain shift. Given a few labeled target domain samples, semisupervised domain adaptation (SSDA) has been explored to enhance the generalization ability of the neural network across domains. However, current SSDA methods focus on domain alignment but fail to achieve fine-grained category-level feature alignment. In this article, we propose a category-level feature alignment network (CFAN), which uses a dual-directional prototype alignment module to achieve category-level feature alignment of the source and target domains. In particular, we utilize pseudo-labeling technique to generate pseudo labels for unlabeled target domain features and build target domain prototypes. Nonetheless, the model performance suffers from the noise conveyed in the pseudo labels. To tackle this issue, we further employ the adversarial training strategy to align the multimodal information of different domains and thus improve the quality of pseudo labels. The proposed CFAN outperforms state-of-the-art (SOTA) methods on commonly used remote sensing datasets, with respective mean accuracies of 92.9%, 93.3%, and 94.3% with EfficientNet_B0, VGG16, and ResNet34 backbones in12 adaptation scenarios.