Research on Few-Shot Remote Sensing Scene Classification Based on Residual Network and Transformer

Aiming at the problems that the amount of labeled sample data of remote sensing images is small, which is not suitable for traditional deep learning algorithms, and that remote sensing images have rich spatial semantic features and large intra-class amplitude differences, a few-shot remote sensing scene classification model RN_CST based on the residual network and Transformer is constructed. This model is used for the few-shot remote sensing image scene classification task in the field of national defense and military industry, and it is of great significance in aspects such as military reconnaissance, target recognition and positioning. A learnable class prototype representation is utilized to replace the traditional average prototype representation, and a Matrix Cosine Similarity Attention (MCSA) is proposed to replace the SoftMax Attention (SMA) in the Transformer, so as to reduce the influence of the amplitude differences between two different input feature sets and output a more stable attention map. The experimental results show that when the experiment is set as 5way-1shot and 5way-5shot, the average classification accuracies of the RN_CST model on the NWPU-RESISC45 and AID datasets are 73.36 ± 0.92%, 86.34±0.48% and 62.27±0.57%, 76.19±0.53% respectively, and the classification effect has been significantly improved.