直推式遥感图像场景零样本分类算法

针对遥感图像场景零样本分类算法中的空间类结构不一致以及域偏移问题，提出基于Sammon嵌入和谱聚类方法结合的直推式遥感图像场景零样本分类算法。首先，基于Sammon嵌入算法修正语义特征空间类原型表示，使其与视觉特征空间类原型结构对齐；其次，借助结构迁移方法得到视觉特征空间测试类原型表示；最后，针对域偏移问题，采用谱聚类方法修正视觉特征空间测试类原型，以适应测试类样本分布特点，提高场景零样本分类准确度。在两个遥感场景集（UCM和AID）上分别获得52.89%和55.93%的最高总体分类准确度，均显著优于对比方法。实验结果表明，通过显著降低视觉特征空间和语义特征空间的场景类别结构不一致性，同时减轻了域偏移问题，可实现语义特征空间类结构知识到视觉特征空间的有效迁移，大幅提升遥感场景零样本分类的准确度。

Transductive zero-shot classification algorithm for remote sensing image scenes

Aiming at solving the structure inconsistency between visual feature space and semantic feature space and domain shift of remote sensing image scenes, this paper proposed a transductive zero-shot classification algorithm for remote sensing image scenes based on Sammon embedding and spectral clustering. Firstly, it modified semantic feature space class prototypes by Sammon embedding to align with the visual feature space class prototypes. Secondly, the algorithm obtained the unseen prototypes in visual feature space via structure transfer. Finally, it modified the unseen class prototypes in visual feature space to adapt to the distribution characteristics of unseen class samples, and to improve the zero-shot classification accuracy of remote sensing image scenes. This algorithm obtains the best overall accuracies of 52.89% and 55.93% on two remote sensing image scene datasets(UCM and AID), respectively, which outperforms the comparative methods. The experimental results show that this algorithm can significantly reduce the class structure inconsistency problem and the domain shift problem. It also makes the effective transfer for semantic space class structure knowledge into visual space, and largely improves the accuracies of zero-shot classification for remote sensing image scenes.