Super-Resolution Reconstruction of Satellite Images Using Refined SwinIR

With the explosion in the number of low-cost satellites, satellite imagery has become widely used for a variety of applications, including agriculture, landscape, and environmental awareness. Therefore, one of the immediate challenges is to increase the resolution of images to extract useful information. Accordingly, higher resolution satellite images are expected to improve recognition accuracy. On the other hand, there are several possible limitations to the use of satellite imagery. Therefore, there is a need to develop a method that can increase the resolution of images from as little satellite image data as possible. In this study, we developed a deep learning model with a small number of parameters. We developed Refined SwinIR, an improved version of the conventional method that reduces the number of parameters while maintaining the quality of the results. Satellite images from Google Earth Pro were used for training and testing. Compared to methods based on classical convolutional neural network modules, the number of parameters used in our model was reduced by 44.45%, and we were also able to improve the peak signal-to-noise ratio and structural similarity values.