Few-Shot Remote Sensing Image Scene Classification Model via Multi-Scale Relation Distillation

Remote sensing image scenes classification holds significant application in fields such as disaster response and urban planning. When using deep learning algorithms for scene classification, the limited sample data of remote sensing affect the classification accuracy and stability. A few-shot remote sensing image scene classification method via multi-scale relation distillation is proposed in this paper. First, a multi-scale feature extraction module based on Efficient Pyramid Squeeze Attention Net (EPSANet) is designed to extract the features of the input samples in different scales spaces. Second, a multi-scale relation distillation module based on knowledge distillation (KD) is developed to realize the optimum learning. The cosine similarity function is utilized to compute the relational distribution between the query set samples and the prototype vectors in different scales, and the scale with the largest cosine similarity is selected as the optimal scale. The Kullback–Leibler (KL) divergence between the remaining and the optimal scale relational distributions is then calculated. By aligning the remaining scales with the optimal scale and minimizing the KL divergence, the correlation among multi-scale features is enhanced. The experimental results on RESISC45, UC Merced and AID datasets show that the proposed method consistently improves classification accuracy and outperforms existing approaches.