Improvement and Application of GAN Models for Time Series Image Prediction—A Case Study of Time Series Satellite Nephograms

Abstract Predicting the shape evolution and movement of nephograms is a difficult task required for the effective monitoring and rapid prediction of thunderstorms, gales, rainstorms, and other disastrous weather conditions. With time sequence information as a constraint, forecasting and outputting nephogram data in the future can effectively compensate for the deficiencies of traditional nephogram methods. To improve the visual effect as much as possible and ensure the accuracy of the shape and movement trends of time series nephogram predictions, a GAN (Generative Adversarial Network) is used. Through comparative experiments and analyses, the best combination of the Mish activation function and Huber loss function is selected. We propose an improved GAN model for time series nephogram prediction in the following 3 aspects. A multineighborhood gradient difference loss is adopted to improve the sharpness of the prediction nephogram, a multiscale structure is used to improve the long-term performance of nephogram prediction, and the Wasserstein distance is introduced to replace JS divergence to ensure the normal updating of generator parameters. Experimental Results show that the improved multiscale model can maintain good visual effects and accurately depict the overall changes in and movement trends of nephograms. For example, for the improved model using four-neighborhood gradient difference loss, the overall average improvement in PSNR (peak signal-to-noise ratio) is 6.42% and SSIM (structural similarity) is 35.57%.