Remote Sensing Image Few-shot Incremental Classification Method Based on Margin Constraints and Dual-branch Distillation

In remote sensing applications, models must not only rapidly learn new class knowledge but also retain the ability to recognize old class data, especially in the presence of complex and diverse target categories and limited labeled information. Few-shot incremental learning aims to learn new knowledge with minimal new class samples while preventing the forgetting of old class knowledge. However, overfitting to new class samples and catastrophic forgetting of old class knowledge remain significant issues. To address these challenges, a few-shot incremental classification method based on margin constraints and dual-branch distillation is proposed. By imposing margin constraints between sample categories, the model's ability to recognize base class samples is enhanced, and its generalization performance on new class samples is improved. A graph network is constructed to propagate category information between base and incremental classes, preventing overfitting. Meanwhile, the dual-branch distillation structure reduces the forgetting of old class knowledge, improving model stability. Experimental results demonstrate that the proposed method effectively enhances the retention of old class knowledge and the learning of new class knowledge, significantly improving classification performance.