 Global context encoders, GCEs, are used to extract the contextual information from aerial images, which can help to reduce the impact of adversarial examples on semantic segmentation models. Additionally, global coordinate attention mechanisms, CAMS, are introduced to further enhance the global feature representations by suppressing adversarial noise. Furthermore, a universal adversarial training strategy, UATS, is proposed to improve the robustness of the semantic segmentation model against adversarial example attacks. The results show that GFA net outperforms existing state-of-the-art semantic segmentation models in terms of accuracy and robustness against adversarial examples. This article was authored by Zhen Wang, Bu Hong Wang, Yao Huilu, and others.