Study on Rainfall Threshold Setting for Deep Convolutional Generative Adversarial Networks in Simulating Heavy Rainfall over the Korean Peninsula

To effectively respond to the increasing damage caused by extreme rainfall due to climate change, it is crucial to accurately estimate the rainfall threshold that triggers heavy rain advisories and warnings. This study aims to simulate extreme rainfall patterns associated with monsoons and typhoons over the Korean Peninsula by using daily cumulative rainfall data with a 5km spatial resolution. First, a maximum threshold was determined to represent the characteristics of extreme rainfall data best and a Deep Convolutional Generative Adversarial Network (DCGAN) model was proposed to simulate such extreme rainfall events. The quality of the generated rainfall images was evaluated using metrics such as Frechet Inception Distance (FID), reconstruction loss, and Mean Absolute Percentage Error (MAPE). Analysis results indicated that setting the threshold at the top 0.01% of rainfall events achieved the most effective trade-off in preserving information. The DCGAN model proved useful for quantitative rainfall estimation in monsoon and typhoon scenarios, showing superior performance across all metrics, especially for events with rainfall amounts below 200mm/day, compared to conventional models. The generated images effectively captured localized rainfall patterns and are expected to provide valuable information for predicting and preparing for potential natural disasters in specific areas.