RS-SAGAN: Boosting Remote Sensing Sample Diversity via Seasonal Adaptive GAN

Huan Zhou, Qingyu Liu, Mi Zhang, Yuanxin Zhao, Jiangyong Ying, Yueran Qi

Abstract: While remote sensing training sample volume is increasing, the seasonal characteristics are still lacking despite their crucial role in intelligent interpretation. We address this issue via the remote sensing seasonal adaptive generative adversarial network (RS-SAGAN). RS-SAGAN is composed of a seasonal domain adaptive learning module and a season transfer module. First, the seasonal domain adaptive learning module unifies the data distributions of the source and target domains into the same feature space. Second, the season transfer module converts the original samples to different seasons through a generative adversarial network (GAN). Using Postdam, and Vaihingen datasets as semantic samples, we achieve the competitive Fréchet Inception Distance (FID) score by approximately 5% over the existing approaches. The report on the SEN1-2 dataset with seasonal information demonstrates that RS-SAGAN can eliminate the domain gap and obtain high-quality seasonal samples, enhancing the generalization and robustness of remote sensing interpretation.