Remote Sensing Scene Image Classification with Swin Transformer-Based Transfer Learning

In remote sensing classification problems, high visual similarity between scenes reduces the classification performance of traditional methods. Therefore, advanced deep neural network models are preferred over traditional methods for classifying these images. In this study, the classification performance of the Swin Transformer model, a deep learning method that has attracted considerable attention in recent years, is comprehensively evaluated on two datasets widely used in remote sensing and compared with popular studies in the literature. Thanks to the hierarchical and window-based self-attention mechanism of the Swin Transformer model, the model effectively learns both global and local contextual information from the images and improves the scene classification performance of the model. In order to obtain reliable and generalisable results, 5 -fold cross-validation method was applied to the dataset. The results showed that the Swin Transformer model achieved impressive results with $\mathrm{9 6. 5 2 \%}$ and $\mathrm{9 5. 3 9 \%}$ accuracy rates, outperforming advanced models such as classical CNN, Ensemble and GAN-based models.