Yoon et al. (2026) A heterogeneous weighting strategy for leveraging Cross-Basin data enhances the Usability of deep learning hydrological models

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-02-05
• Authors: Sunghyun Yoon, Dongkyun Kim, Kuk-Hyun Ahn
• DOI: 10.1016/j.jhydrol.2026.135097

Research Groups

• Department of Artificial Intelligence, Kongju National University, Cheon-an, South Korea
• Department of Civil and Environmental Engineering, Hongik University, Seoul, the Republic of Korea
• Department of Civil and Environmental Engineering, Kongju National University, Cheon-an, South Korea

Short Summary

This study develops a novel heterogeneous weighting strategy for deep learning hydrological models to effectively leverage cross-basin data, demonstrating improved predictive performance over conventional homogeneous weighting by accounting for basin-specific characteristics. The proposed method enhances the usability of deep learning models for hydrological prediction by mitigating local performance degradation often seen in regional pooling models.

Objective

• To develop and evaluate a novel heterogeneous weighting strategy for deep learning hydrological models (specifically LSTM) to enhance their usability and predictive performance when leveraging cross-basin data, particularly for optimizing training for individual basins while avoiding local performance degradation.

Study Configuration

• Spatial Scale: 1,678 basins in total, comprising 531 basins from the CAMELS-US dataset (United States) and 1,147 basins from the CAMELS-DE dataset (Germany).
• Temporal Scale: Not explicitly stated, but typical for hydrological modeling studies using CAMELS datasets, often spanning several decades.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) networks, a type of deep learning (DL) model.
• Data sources: CAMELS-US dataset and CAMELS-DE dataset.

Main Results

• Regional pooling models frequently suffer from local performance degradation when indiscriminately using all available cross-basin data.
• The proposed heterogeneous weighting method delivers improved predictions compared to conventional homogeneous weighting strategies.
• The approach achieved a 0.028 increase in median Kling–Gupta Efficiency compared to the homogeneous weighting strategy, a difference statistically significant at the 99.9% confidence level.
• Carefully designed training strategies, such as the proposed heterogeneous weighting, provide greater gains than simply adding additional input variables to the network when modeling over 1,678 basins across two countries.

Contributions

• Introduces a novel heterogeneous weighting strategy that quantifies the influence of one basin’s gradient on the loss of a target basin, offering insights into inter-basin relationships.
• Optimizes deep learning model training for individual basins by explicitly accounting for basin-specific characteristics, thereby enhancing predictive capabilities at both local and global scales.
• Demonstrates that a sophisticated training strategy is more effective for improving hydrological prediction performance than merely increasing the number of input variables.
• Advances the usability and reliability of deep learning hydrological models by effectively leveraging large cross-basin datasets while mitigating local performance degradation.

Funding

• Not explicitly mentioned in the provided text.

Citation

@article{Yoon2026heterogeneous,
  author = {Yoon, Sunghyun and Kim, Dongkyun and Ahn, Kuk-Hyun},
  title = {A heterogeneous weighting strategy for leveraging Cross-Basin data enhances the Usability of deep learning hydrological models},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135097},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135097}
}