Year: 2025
Type: conference-paper
Abstract: The intensification of extreme rainfall events due to climate change has significantly increased flood risks, with levee overtopping emerging as a major cause of catastrophic failures.Accurate short-term water level forecasting is therefore essential, particularly in ungauged river systems where monitoring infrastructure is limited.However, existing deep learning models heavily rely on long-term observations from gauged sites, while physics-based models require dense, high-quality input data-both of which constrain real-time applicability in data-scarce regions.To address these challenges, we propose DeepCreek, a lightweight MLP-based forecasting model adapted from TSMixer, designed for short-term water level forecasting in ungauged basins.Utilizing nationwide rainfall scenario-based datasets comprising eight key hydrometeorological variables across 232 observation stations in South Korea, and employing a spatially segregated evaluation strategy, DeepCreek achieved an RMSE of 1.0987m and a MAPE of 15.43% for six-hour-ahead predictions under extreme rainfall scenarios.The model consistently outperformed conventional approaches on both the rainfall scenario and extreme rainfall scenario test sets, demonstrating robust performance even under noisy real-world conditions without explicit outlier removal.These results underscore DeepCreek's potential as a practical and scalable AI solution for real-time flood forecasting in regions with limited hydrological monitoring infrastructure.
Authors: Younga Shin, Kichang Choi, Yeonjoo Kim, Hongjo Kim
Institution: Yonsei University
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: bronze
Awards: G7517803420, G8811246431
OpenAlex ID: W4413220846