CZSRSSC: Continual Zero-Shot Remote Sensing Scene Classification

Remote sensing scene classification, a core technology in fields such as disaster response, resource management, and urban planning, often faces challenges in real-world applications, such as identifying unlabeled new classes (zero-shot learning) and dynamically expanding classes (continuous learning). While continual zero-shot scene classification is actively researched in common scenarios, it remains largely unresolved in remote sensing scenarios. Furthermore, a major issue currently hindering the continued development of zero-shot scene classification in remote sensing is the lack of specialized datasets. To address these issues, we conducted the following research: First, we introduced continual zero-shot scene classification into the remote sensing field for the first time, breaking the traditional remote sensing scene classification’s reliance on predefined classes of all classes and achieving effective classification of dynamically increasing remote sensing scene classes. Secondly, we established the first dedicated remote sensing dataset for continual zero-shot scene classification tasks by integrating multiple public remote sensing datasets, constructing unified semantic information, selecting high-quality samples, and extracting image and attribute features. Finally, we verified the effectiveness of the existing continual zero-shot scene classification method in common scenes in remote sensing scenarios based on the RS-CZSL dataset, and compared it with the common zero-shot scene classification method.