Ougahi et al. (2026) Investigating Deep Learning Knowledge Transfer in Streamflow Prediction From Global to Local Catchment

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

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-02-01
• Authors: Jamal Hassan Ougahi, John S. Rowan
• DOI: 10.1029/2025wr041194

Research Groups

Not explicitly available in the provided abstract. The study involves data from Scotland (GB-SCT), Switzerland (CH), British Columbia (BC), and California (CA).

Short Summary

This study evaluates transfer learning approaches using Long Short-Term Memory (LSTM) models to improve streamflow prediction in data-scarce regions by leveraging data from data-rich areas and catchment characteristics. It demonstrates that pre-training LSTMs on hydrologically similar basin clusters, followed by fine-tuning with limited local data, significantly enhances prediction accuracy in data-poor regions.

Objective

• To evaluate the effectiveness of transfer learning (TL) approaches using Long Short-Term Memory (LSTM) models for streamflow prediction in data-scarce regions, by pre-training on data-rich donor basins and fine-tuning on data-poor target basins.

Study Configuration

• Spatial Scale: 441 donor basins across Scotland, Switzerland, and British Columbia; 36 target basins in California.
• Temporal Scale: Time-series streamflow records (lagged) and global climate data (ERA-5). Specific duration not specified in abstract.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) deep learning models, Transfer Learning (TL) techniques, K-Means clustering algorithm.
• Data sources: Measured streamflow records, global climate data (ERA-5), catchment attributes.

Main Results

• Transfer learning (TL-LSTM) models performed better than locally trained models, achieving a Nash-Sutcliffe Efficiency (NSE) of 0.85 and a Kling-Gupta Efficiency (KGE) of 0.80.
• The LSTM model pre-trained on basins from Cluster 3, which most closely resembled the target region's hydrology, yielded the most accurate predictions.
• Fine-tuning with even limited local data substantially improved prediction accuracy in validation splits (treated as ungauged basins).
• Pairing measured streamflow records (lagged) with global climate data (ERA-5) significantly boosted the explanatory power of LSTM predictions, particularly in snowmelt and glacier-influenced basins.

Contributions

• Demonstrates the effective application of TL-LSTM for enhancing streamflow prediction in data-scarce hydrological regions.
• Advances the understanding of cross-basin generalization capabilities of deep learning models in hydrology.
• Supports the development of efficient and scalable modeling strategies for hydrological prediction in areas with limited observational data.

Funding

Not available in the provided abstract.

Citation

@article{Ougahi2026Investigating,
  author = {Ougahi, Jamal Hassan and Rowan, John S.},
  title = {Investigating Deep Learning Knowledge Transfer in Streamflow Prediction From Global to Local Catchment},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041194},
  url = {https://doi.org/10.1029/2025wr041194}
}