Analyzing the predictive power of GAN-based precipitation prediction model against a process-based model: a case study of the Yeongsangang River basin in South Korea

Radar-based methods or physics-based Numerical Prediction Models (NWPs) have been used to generate high-resolution short-range rainfall forecasts. With recent advances in artificial intelligence, research on radar-based short-range precipitation nowcasts using machine learning has been actively conducted and has shown promising performance for short-range forecasts within two hours. However, these data-driven models have the limitation of being the black-box models that do not consider atmospheric physical processes, and their performance decreases significantly as the forecast lead time increases. In this study, we developed a precipitation nowcasting model using the conditional Generative Adversarial Network (cGAN) for the Yeongsangang River basin, which generates predictions up to six hours in advance, and evaluated its performance through a comparison with Weather Research and Forecasting (WRF), a physical process-based atmospheric model capable of high-resolution numerical simulations. To train and evaluate the models, we used rainfall radar data from the Yeongsangang River basin from 2014 to 2018. During the three-hour lead time, the cGAN-based rainfall forecast model showed 10% and 16% better performance in the average correlation coefficient R and the critical success index (CSI, at 0.1 mm/hr) than the WRF. However, as the lead time increased after three hours, the WRF model's forecast performance was 55% and 39% better than the cGAN-based model in the R and CSI (at 0.1 mm/hr), respectively. In this study, we explored the strengths and weaknesses of machine learning-based models and physics-based models, suggesting possible opportunities to complement them. It is expected that our results will provide a foundation for a physics-AI integrated approach to enhance rainfall forecast performance.