A Novel Remote Sensing Benchmark for Few-Shot Class-Incremental Learning Based on NWPU-RESISC45

In the field of remote sensing, scenes are inherently dynamic, with new observation targets continuously emerging, requiring analysis models to consistently adapt to novel classes. Meanwhile, the acquisition of remote sensing data is challenging, and the annotation process is costly, resulting in a scarcity of labeled samples. This issue is further exacerbated when new targets appear with extremely limited labeled data. Under these circumstances, Few-Shot Class-Incremental Learning (FSCIL), which is dedicated to learning new knowledge from novel classes with only a few samples while retaining previously acquired knowledge, becomes particularly significant for remote sensing. However, existing research in this area remains limited. To bridge this gap, our work constructs a new benchmark for FSCIL in remote sensing based on the public NWPU-RESISC45 dataset, and further develops an experimental protocol and provides baseline results for this benchmark. We evaluate five representative FSCIL methods on this benchmark and provide a comprehensive analysis of their performance. Experiments in our paper demonstrate the feasibility and effectiveness of our benchmark and establish a solid foundation for advancing FSCIL techniques in the remote sensing field.