DYNAMIC LEARNING RATE FOR FEW-SHOT INCREMENTAL LEARNING IN OPTICAL REMOTE SENSING IMAGE CLASSIFICATION

Few-shot incremental learning is crucial for optical remote sensing images. It enables the model to learn the classification of new classes from a limited number of labeled samples without forgetting previously learned classes. The classical few-shot incremental learning methods include training the backbone network and fine-tuning the classifier. However, existing methods apply the same learning rate to feature vectors across all classes during the fine-tuning process, which leads to the forgetting of previously learned old classes, thereby degrading classification performance. To address this issue, we propose a dynamic learning rate mechanism. Specifically, different learning rates are assigned to the feature vectors of new and old classes respectively, so that the feature vectors of new classes can iterate normally while reducing the iteration speed of the feature vectors of old classes. This differential treatment allows the model to adaptively focus on learning new knowledge while safeguarding the old knowledge. Additionally, we introduce knowledge distillation techniques to enhance the feature extraction capability of the backbone model. Extensive experiments on public optical remote sensing classification datasets demonstrate that our method gets superior performance compared with several state-of-the-art methods.