Adversarial Self-Training Unsupervised Domain Adaptation for Remote Sensing Scene Classification
Ying Chun Huang, Tangsheng Li, Can Liu, Wenhao Mei · IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium · 2022
In recent years, deep learning have achieved outstanding performance in field of remote sensing scenes classification. At the same time, most neural network models also exposed many shortcomings when dealing with different imaging conditions, insufficient sample size, etc. By analyzing the characteristics of source domain and target domain, reducing the gap between these domains is the main key for better recognition performance. In this paper, a novel adversarial self-training unsupervised domain adaptation (AST) framework is proposed to deal with the domain migration issues. To enforcing the gap reducing and the alignment of feature distribution of domains, we first implement the domain-adversarial learning mechanism. And then the self-training modules is also applied for better decision boundary generation. The experimental results show that our proposed method outperforms state of art in remote sensing scene classification.