Zero-shot remote sensing image scene classification based on robust cross-domain mapping and gradual refinement of semantic space

• Main Authors: LI Yansheng, KONG Deyu, ZHANG Yongjun, JI Zheng, XIAO Rui
• Format: Article
• Language: zho
• Published: Surveying and Mapping Press 2020-12-01
• Series: Acta Geodaetica et Cartographica Sinica
• Subjects: zero-shot learning; remote sensing image scene classification; cross-domain mapping with auto-encoder; collaborative representation learning; natural language processing
• Online Access: http://xb.sinomaps.com/article/2020/1001-1595/2020-12-1564.htm

Zero-shot classification technology aims to acquire the ability to identify categories that do not appear in the training stage (unseen classes) by learning some categories of the data set (seen classes), which has important practical significance in the era of remote sensing big data. Until now, the zero-shot classification methods in remote sensing field pay little attention to the semantic space optimization after mapping, which results in poor classification performance. Based on this consideration, this paper proposed a zero shot remote sensing image scene classification method based on cross-domain mapping with auto-encoder and collaborative representation learning. In the supervised learning module, based on the class semantic vector of seen class and the scene image sample, the depth feature extractor learning and robust mapping from visual space to semantic space are realized. In the unsupervised learning stage, based on the class semantic vectors of all classes and the unseen remote sensing image samples, collaborative representation learning and <i>k</i>-nearest neighbor algorithm are used to modify the semantic vectors of unseen classes, so as to alleviate the problem of the shift of seen class semantic space and unseen class semantic space one after another and unseen after self coding cross domain mapping model mapping the shift of class semantic space and unseen class semantic space after collaborative representation. In the testing phase, based on the depth feature extractor, self coding cross domain mapping model and modified unseen class semantic vector, the classification of unseen class remote sensing image scene can be realized. We integrate a number of open remote sensing image scene data sets and build a new remote sensing image scene data set, experiments were conducted using this dataset The experimental results show that the algorithm proposed in this paper were significantly better than the existing zero shot classification method in the case of a variety of seen and unseen classes.