Waqas et al. (2025) Hybrid Deep Learning Versus Empirical Methods for Daily Potential Evapotranspiration Estimation in the Nakdong River Basin, South Korea

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water
• Year: 2025
• Date: 2025-12-22
• Authors: Muhammad Waqas, S. M. Kim
• DOI: 10.3390/w18010032

Research Groups

Researchers focused on the Nakdong River Basin, South Korea.

Short Summary

This study compared empirical and deep learning models for daily potential evapotranspiration (PET) estimation in the Nakdong River Basin, South Korea, finding that a hybrid Convolutional Neural Network Bidirectional LSTM with an attention mechanism significantly outperformed empirical and standalone deep learning models, demonstrating high accuracy and suitability for regional hydrological applications.

Objective

• To compare the performance of empirical and hybrid deep learning (DL) models in estimating daily potential evapotranspiration (PET) in the Nakdong River Basin (NRB), South Korea, using the FAO-56 Penman–Monteith (PM) method as a reference.

Study Configuration

• Spatial Scale: Nakdong River Basin (NRB), South Korea, utilizing data from 13 meteorological stations.
• Temporal Scale: Daily potential evapotranspiration (PET) over a meteorological dataset spanning from 1973 to 2024.

Methodology and Data

• Models used:
  • Empirical models: Priestley–Taylor (P-T), Hargreaves–Samani (H-S).
  • Deep Learning (DL) models: Standalone Long Short-Term Memory (LSTM), Hybrid Convolutional Neural Network Bidirectional LSTM with an attention mechanism (CNN-BiLSTM Attention Mechanism).
  • Reference method: FAO-56 Penman–Monteith (PM).
• Data sources: Meteorological dataset collected from 13 meteorological stations within the Nakdong River Basin.

Main Results

• Empirical models (P-T, H-S) performed poorly, exhibiting a basin-wide Root Mean Square Error (RMSE) of 5.04–5.79 mm/day and negative Nash-Sutcliffe Efficiency (NSE) values ranging from −10.37 to −13.99, indicating their unsuitability for the NRB.
• Deep learning models achieved significant improvements in accuracy. The hybrid CNN-BiLSTM Attention Mechanism (using input combination C1) demonstrated the highest performance with a coefficient of determination (R²) of 0.820, RMSE of 0.672 mm/day, NSE of 0.820, and Kling-Gupta Efficiency (KGE) of 0.880.
• The standalone LSTM model showed lower performance compared to the hybrid model, with R² = 0.756 and RMSE = 0.782 mm/day.
• Spatial analysis confirmed the generalization capability of the hybrid DL model across heterogeneous climates, with station-level NSE consistently exceeding 0.70.
• The hybrid DL model accurately represented the temporal variability and seasonal patterns of PET, making it highly suitable for operational hydrological modeling and water-resource planning in the NRB.

Contributions

• Demonstrated the superior performance of a hybrid deep learning model (CNN-BiLSTM Attention Mechanism) for daily PET estimation in the Nakdong River Basin compared to traditional empirical models and standalone LSTM networks.
• Provided a robust and accurate method for PET estimation in the region, verifying its generalization across heterogeneous climates.
• Highlighted the potential of advanced deep learning architectures for improving operational hydrological modeling and water-resource planning in complex river basins.

Funding

Not specified in the provided text.

Citation

@article{Waqas2025Hybrid,
  author = {Waqas, Muhammad and Kim, S. M.},
  title = {Hybrid Deep Learning Versus Empirical Methods for Daily Potential Evapotranspiration Estimation in the Nakdong River Basin, South Korea},
  journal = {Water},
  year = {2025},
  doi = {10.3390/w18010032},
  url = {https://doi.org/10.3390/w18010032}
}