Jahangir et al. (2026) A novel hybrid fine-tuning method for supercharging deep learning model development for hydrological prediction

Identification

• Journal: Environmental Modelling & Software
• Year: 2026
• Date: 2026-04-01
• Authors: M. S. Jahangir, John Quilty, C. Shen, Andrea Scott, Scott Steinschneider, J. Adamowski
• DOI: 10.1016/j.envsoft.2026.106978

Research Groups

• Department of Bioresource Engineering, McGill University, Montreal, Canada
• Department of Biological and Environmental Engineering, Cornell University, Ithaca, United States
• Department of Civil and Environmental Engineering, University of Waterloo, Waterloo, Canada
• Department of Civil and Environmental Engineering, Pennsylvania State University, College town, United States
• Department of Mechanical Engineering, University of Waterloo, Waterloo, Canada
• United Nations University, Institute for Water, Environment and Health, Richmond Hill, Canada

Short Summary

This study introduces a novel hybrid Long Short-Term Memory (LSTM) and Random Forest (RF) fine-tuning method that significantly accelerates and enhances deep learning model development for streamflow prediction, demonstrating superior efficiency and accuracy compared to conventional methods.

Objective

• To propose and evaluate a novel hybrid Long Short-Term Memory (LSTM) and Random Forest (RF) transfer learning method to reduce computational costs, improve accessibility, and enhance the performance of deep learning models for streamflow prediction, particularly in data-limited or computationally constrained scenarios.

Study Configuration

• Spatial Scale: 421 catchments in the continental United States (CAMELS-US) and 324 catchments in Germany (CAMELS-DE). Catchment areas in CAMELS-DE range from 5 km² to 15,000 km².
• Temporal Scale: Daily time step.
  • CAMELS-US: Data from June 1984 to December 2012 (approximately 28.5 years). Training/validation period: June 6, 1984, to November 24, 2007. Testing period: November 25, 2007, to January 15, 2012.
  • CAMELS-DE: Data from January 1951 to December 2020 (up to 70 years, median 46 years). Training period: October 1, 1970, to December 31, 1999. Validation period: October 1, 1965, to September 30, 1970. Testing period: January 1, 2001, to December 31, 2020.
  • A lookback period of 365 days was adopted for all experiments.

Methodology and Data

• Models used:
  • Hybrid Long Short-Term Memory (LSTM) encoder with a Random Forest (RF) decoder.
  • Regional LSTM models (with varying numbers of catchments and static features).
  • Catchment-wise LSTM models.
  • LSTMA (fine-tuning all LSTM layers).
  • LSTMB (fine-tuning only the linear head of LSTM).
  • Adam optimizer for LSTM training.
• Data sources:
  • CAMELS-US dataset (Catchment Attributes and MEteorology for Large-sample Studies - United States).
  • CAMELS-DE dataset (Catchment Attributes and MEteorology for Large-sample Studies - Germany).
  • Daymet meteorological variables for CAMELS-US: precipitation (mm/day), maximum and minimum air temperature (°C), shortwave downward radiation (W/m²), and vapor pressure (Pa).
  • CAMELS-DE input features: mean precipitation (mm/day), precipitation standard deviation (mm/day), mean radiation (W/m²), and mean minimum and maximum temperature (°C).
  • Specific discharge (mm/day) as the target variable, sourced from Caravan for CAMELS-US.
  • Static catchment attributes: 4 features for Scenario 1, and 27 features for Scenario 2.

Main Results

• The hybrid method significantly outperformed other transfer learning (TL) strategies (LSTMA and LSTMB) in Scenario 1 (limited source data), showing a 3.99% (KGE) and 4.84% (NSE) median improvement over LSTMA-50, and 1.11% (KGE) and 6.04% (NSE) over LSTMB-50 for CAMELS-US.
• The hybrid-50 model achieved 4.14% (KGE) and 9.19% (NSE) median improvements compared to catchment-wise training.
• The hybrid method was computationally more efficient, being approximately 35% faster than LSTMA-50 and 32% faster than LSTMB-50 in optimization time.
• For CAMELS-DE, the hybrid-50 model performed competitively with a benchmark regional model (trained on 1555 catchments), achieving a minimum Nash-Sutcliffe Efficiency (NSE) of 0.369 compared to 0.094 for the benchmark.
• In Scenario 2 (ample source data), the hybrid approach (hybrid-371) led to substantial performance improvements for out-of-sample catchments (14.11% median KGE and 13.13% median NSE improvement) compared to the regional LSTM-371 model.
• Fine-tuning a sub-optimal regional model (hybrid-421) was more beneficial than fine-tuning an optimal regional model trained on fewer catchments, resulting in median KGE and NSE gains of 0.92% and 1.19%, respectively, over the optimal regional LSTM-421.
• Statistical analysis (Wilcoxon test) revealed that KGE significantly increased in 53.51% of catchments and NSE significantly increased in 43.60% of catchments when comparing hybrid-421 to the regional LSTM-421.
• The hybrid model reduced the prevalence of basins with negative NSE values (21 out of 421) compared to the regional LSTM-421 model (31 out of 421).
• The hybrid approach improved performance in catchments where the regional model initially underperformed (KGE and NSE below 0.75), increasing median KGE from 0.762 to 0.788 and NSE from 0.758 to 0.773.

Contributions

• Proposes a novel hybrid transfer learning approach that integrates a regional Long Short-Term Memory (LSTM) encoder with a Random Forest (RF) decoder for efficient and accessible fine-tuning of deep learning models in hydrology.
• Demonstrates that this hybrid method offers superior performance and computational efficiency compared to conventional LSTM fine-tuning strategies (LSTMA and LSTMB).
• Shows that applying fine-tuning to a regional model, even if not fully optimized or trained on limited data, yields better predictive accuracy than relying solely on regional modeling or catchment-wise training.
• Provides a practical and accessible solution for hydrologists and water resource practitioners to leverage advanced regional deep learning models for site-specific applications without requiring extensive deep learning expertise or high-performance computing resources.
• Evaluates the hybrid approach as a complementary tool for large-sample regional streamflow prediction, particularly in improving performance for out-of-sample catchments and those where regional models underperform.

Funding

• Natural Resources Canada (PHIMP23-27P3)

Citation

@article{Jahangir2026novel,
  author = {Jahangir, M. S. and Quilty, John and Shen, C. and Scott, Andrea and Steinschneider, Scott and Adamowski, J.},
  title = {A novel hybrid fine-tuning method for supercharging deep learning model development for hydrological prediction},
  journal = {Environmental Modelling & Software},
  year = {2026},
  doi = {10.1016/j.envsoft.2026.106978},
  url = {https://doi.org/10.1016/j.envsoft.2026.106978}
}