Very Short-term Prediction of Weather Radar-Based Rainfall Distribution and Intensity Over the Korean Peninsula Using Convolutional Long Short-Term Memory Network

Rain is one of the most important atmospheric phenomena that is directly related to ecosystems and human lives. Numerous meteorological organizations operate weather radars to measure in-situ rainfall. This study performs very short-term prediction of rainfall distribution and intensity for up to 2.5 h over the Korean Peninsula experiencing various rain types, such as typhoon, stationary-front, and localized heavy rainfalls, using convolutional Long Short-Term Memory (ConvLSTM) networks and the constant altitude plan position indicator data provided by the Korean Meteorological Administration (KMA). The ConvLSTM-based prediction models were trained using 3522 cases of radar-based rainfall data from April to October in 2018 and 2019, as well as 1584 typhoon, stationary-front, and localized heavy rainfall events during August and September 2020, respectively. The rainfall presence prediction model showed a decrease in the critical success index (CSI) (0.8659 to 0.7266, 0.7366 to 0.4821, and 0.5302 to 0.0946 for typhoon, stationary-front, and localized heavy rainfalls, respectively) from 10 min to 2.5 h. The rainfall intensity prediction model showed an Heidke skill score of 0.4741 for light rain (0.1 to 3 mm/h) after 1 h. However, heavy and very heavy rainfall events were not predicted due to the lack of their training datasets and short lifetimes, which was supported by the relationship between the average rain event rate (for 150 min) and CSI index. Consequently, compared to KMA’s operational MAPLE model. The ConvLSTM-based models are effective in predicting rainfall type and its intensity over wide areas for a relatively long time.