Generating High-Resolution Climate Prediction through Generative Adversarial Network

Abstract Downscaling technology is always used for high-resolution climate prediction, and this technology can generate small-scale regional climate prediction from large-scale climate output information. Inspired by image super resolution, we propose to apply the super-resolution models to downscaling technology. However, some unpleasant artifacts always appear in the high-resolution climate images generated by exiting super-resolution models. To further eliminate these unpleasant artifacts, we innovatively apply the super-resolution generative adversarial network (SRGAN) to climate prediction. SRGAN adopts a perceptual loss function which consists of an adversarial loss and a content loss, which can recover more high-frequency details in the generated high-resolution images. Besides, we propose a method to fuse climate data with related meteorological factors to generate climate images, this measure can improve the accuracy of climate prediction. Finally, extensive experimental results on climate datasets show that SRGAN performs better than most super-resolution approaches in climate prediction.