| Title |
A Study on Optimizing Training Data to Improve Accuracy of DeepLearning-based Real-time River Flood Prediction |
| Authors |
윤성심(Yoon, Seong-sim);최지안(Choi, Gian) |
| DOI |
https://doi.org/10.12652/Ksce.2025.45.3.0347 |
| Keywords |
딥러닝, 실시간, 최적화, 하천, 학습데이터, 홍수예측 Deep learning, Flood prediction, Optimizing, Real-time, River, Training data |
| Abstract |
One of the primary objectives of predicting river water levels is to establish criteria for issuing flood warnings and alerts. This study aims
to optimize the training data for a deep learning-based river water level prediction model and enhance its accuracy by utilizing
AutoKeras, which supports automatic design and optimization of deep learning models, to develop models that minimize artificial
influences. The upper basin of the Hantan River was selected as the study area, and datasets were constructed using water level data
from three observation stations and mean areal rainfall data. Based on these datasets. Based on these datasets, Two models were
developed: Model 1 was trained on datasets that included all recorded rainfall events, while Model 2 was trained on datasets capturing
significant water level increases. Predictions for Hantan Bridge indicated that Model 1 achieved higher accuracy in time-series water
level estimation, as evidenced by a higher correlation and lower RMSE. In contrast, Model 2 exhibited superior flood detection
capability, showing higher recall, F1-score, and CSI. These results highlight the importance of selecting appropriate training data when
developing deep learning models, particularly for flood prediction. Emphasizing critical factors such as water level rises can enhance
model performance, enabling more effective early warning systems and improving disaster preparedness. |