Development of a Deep Learning–Based Rainfall Prediction System for Urban-Scale Hydrological Disaster Response

Over the past 50 years (1973–2022), South Korea has experienced a modest increase in average precipitation. Notably, the maximum hourly rainfall has significantly risen (Kim et al., 2022). There has been an observed intensification in the intensity and frequency of extreme precipitation events, leading to substantial socioeconomic damages, flash floods, and urban flooding with severe consequences (Dave et al., 2021). Addressing locally occurring extreme precipitation in densely populated areas requires accurate urban-scale rainfall predictions. Given the anticipated increase in future precipitation variability on the Korean Peninsula, detailed spatiotemporal prediction technologies are crucial for effectively managing and responding to extreme precipitation events.

We developed a rainfall prediction system through deep learning approach. In this study, we crafted a rainfall prediction system based on U-Net, a deep-learning architecture widely employed as a foundational model in previous rainfall prediction studies (Badrinath et al., 2023; Han et al., 2023; Lyu et al., 2023). The Advantage of U-Net lies in its end-to-end usability, minimizing the need for manual feature extraction, even with limited training data in the weather domain. To predict rainfall patterns over time, we adopted a recursive approach, drawing inspiration from prior research (Ayzel et al., 2020). In the case of the concentrated rainfall event in Osong in South Korea, July 2023, comparing QPE (Quantitative Precipitation Estimation) with the proposed rainfall prediction technology showed superior performance in terms of spatial patterns and rainfall intensity for the 10-minute lead time. Conversely, numerical weather prediction (Korea Local Analysis and Prediction System) failed to capture the rainfall pattern. For the 180-minute lead time, numerical prediction successfully detected rainfall, while the proposed prediction technology did not capture the rainfall pattern.

Despite being in the early stages of development, case studies validate that our proposed system effectively simulates rainfall patterns that traditional nowcasting or numerical methods may not accurately replicate. However, limitations emerged in predicting localized rainfall intensity as the prediction time lengthened, revealing a tendency for spatial patterns of rainfall to be smoothed. As a follow-up study, our objective is to explore the applicability of deep learning across various aspects of the rainfall prediction process. This includes investigating super-resolution and blending of data produced by existing rainfall prediction methods, conducting empirical studies of deep learning models for domestic heavy rainfall cases, and optimizing existing numerical models. Our goal is to assess the feasibility of deep learning and enhance the accuracy of continuous prediction technology.