| Abstrakt: |
Accurate flood water level prediction is crucial for effective mitigation strategies. However, limited data often hinders the development of robust models. This study explores transfer learning (TL) with warmup learning rate scheduling technique to improve flood prediction using Long Short-Term Memory (LSTM) models. The study focused on watersheds in the United States with limited data. A well-trained source LSTM model from a Tennessee watershed served as the knowledge base for transfer to two target Kentucky watersheds and one target Tennessee watershed. In a comparative evaluation with multiple locally trained LSTM models within these target watersheds, the TL models achieved better performance across different flood magnitude events (low, medium, high), while requiring less training time due to leveraging pre-trained knowledge. In the first target watershed, Root Mean Squared Error (RMSE) reduction was 42.72%, 46.37%, and 82.2% for low, medium, and high flood events, respectively. Similarly, RMSE reductions were also observed in the second target watershed (72.72%, 61.7%, and 64.52%) for low, medium, and high magnitude events, respectively. Finally, while the third watershed showed a decline in performance for the high flood event (12.5% RMSE increase), the TL model slightly performed better than the local models for low and medium flood magnitude events (9% and 3% RMSE decrease). In general, this study highlights the effectiveness of TL with warmup learning rate scheduling technique in LSTM models for flood prediction. This approach not only balances computational efficiency with predictive accuracy but also demonstrates reliable performance across different flood magnitude events, as shown by the statistical significance based on Levene's test. |
Full text from SpringerLink