Super Resolution Generative Adversarial Network Model parameter optimization to improve the quality of geostationary meteorological satellite images

The purpose of this study was to determine the "batch size" and "epoch" parameters of the SRGAN (Super Resolution Generative Adversarial Network) Model that best improve the quality of images from the Second Generation Geostationary Meteorological Satellites. The datasets used for training the model consisted of two hundred images from this satellite, including one hundred low-resolution images and one hundred corresponding high-resolution images. The images were captured by the same meteorological satellite at the same moments but in two different resolutions (low and high resolution). 80 pairs of images were used for model training, and the remaining 20 pairs were used for testing. For each combination of batch size and epoch parameters, the trained model generated an image, which was then compared to the expected high-resolution image using the PSNR metric. The PSNR evolution curves as a function of these parameters taken independently (the other parameters of the SRGAN model were set to default) helped find the optimal values for batch size and epoch.