【摘要】 遥感图像场景分类旨在根据遥感图像的内容为其自动赋予相应的语义标签，已成为当前遥感图像处理领域中的研究热点。基于卷积神经网络（Convolutional Neural Networks, CNNs）的方法和基于自注意力机制的方法则是当前遥感图像场景分类中的两大主流方法。然而，前者不擅长学习长程上下文关系；后者对局部信息的学习能力有限，且具有较大的参数量和运算量。针对上述问题，提议一种基于知识蒸馏的轻量化遥感图像场景分类方法。该方法分别以Swin Transformer和小型CNN网络作为教师模型和学生模型，通过知识蒸馏的方式融合两种模型的优势；更进一步，提出一种新颖的知识蒸馏损失函数，使学生模型能够同时关注遥感图像类间和类内的潜在信息。在两个大规模数据集上的实验结果表明，与现有其它方法相比，所提出方法不仅有高的分类精度，还具有显著降低的参数量和运算量。更多还原

【Abstract】 Remote sensing image scene classification aims to automatically assign a semantic label to each remote sensing image according to its content, and has become one of the hot topics in the field of remote sensing image processing. Methods based on convolutional neural networks(CNNs) and methods based on self-attention mechanism are two mainstream methods in remote sensing image scene classification. However, the former is less effective in exploring long-range contextual information, and the latter has limitations in learning local information and has a large number of parameters and calculations. In order to address these issues, a lightweight method based on knowledge distillation is proposed to solve the problem of scene classification for remote sensing images. The proposed method uses Swin Transformer and lightweight CNNs as the teacher model and the student models, respectively, and integrates the advantages of the two kinds of models by means of knowledge distillation. Furthermore, a novel distillation loss function is proposed to enable the student models to focus on both inter-and intra-class potential information of remote sensing images simultaneously. The experimental results on two large-scale remote sensing image datasets demonstrate that the proposed method not only achieves high classification accuracy compared to existing methods but also has a significantly reduced number of parameters and calculations.更多还原