Comparison of Swin-based transformers as preprocessing in remote sensing image classification

In image classification applications, the quality of the input images is a critical factor in improving classification accuracy. Remote sensing images often have quality issues related to low resolution and sharpness, which can pose challenges in the analysis process. In this context, quality enhancement and correction are of great importance for the classification of remote sensing images. In this study, the classification performance of the DenseNet121, Xception, and EfficientNetV2B3 models was compared using Swin-based HST, Swin2SR, SwinIR, and SwinFIR super-resolution image enhancement on the RSI-Cb remote sensing dataset. During the classification studies, Adam, SGD, RMSprop, and Adagrad optimization algorithms were investigated. As a result of the experiments, it was found that the applied super-resolution methods improve the observed classification accuracy. The best result was obtained using the Adam optimization algorithm with the Densenet121 model of the SwinIR super-resolution method, increasing the F1 value from 99.45% to 99.53%. The RSI-Cb remote sensing data set was divided into man-made and natural regions, and better results were obtained for the natural regions. Furthermore, by testing 350 newly acquired images, it was observed that Swin-based super-resolution methods improved the classification. In addition, experiments were conducted on historical site images to identify historical sites as man-made structures. As a result of the experiments, the SwinFIR Xception model with the Adam optimization algorithm identified historic sites as man-made with 100% accuracy. These results demonstrate that the use of Swin-based methods on remote sensing images can significantly improve classification performance.