Class Centralized Dictionary Learning for Few-Shot Remote Sensing Scene Classification

Recently, few-shot scene classification has become an important task in the remote sensing (RS) field, mainly solving how to obtain better classification performance when there are insufficient labeled samples. The few-shot scene classification task includes the pretrain stage and meta-test stage. There is no category intersection between these two stages. Thus, the sample distribution of the training set and meta-test set is different, leading to the training model's weak generalization or portability. To solve this problem, we propose a class-centralized dictionary learning (CCDL) method for the few-shot RS scene classification (FSRSSC). Specifically, in the pretraining stage, we adopt the model pretrained on a large natural images dataset and then fine-tune the network by the RS dataset. Using the pretrained model helps improve the model's generalization ability. In the meta-test stage, we propose a CCDL classifier, which guarantees the sparse representations of different categories more distant and the same more concentrated. We experiment on several benchmark datasets and achieve superior performance, demonstrating the proposed method's effectiveness.