Sezen et al. (2026) Robust discharge prediction of seasonal snow-influenced karst systems through hybridization of process-based and data-driven models

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-01-22
• Authors: C. Sezen, N. Ravbar, Andreas Hartmann, Kai Chen
• DOI: 10.1016/j.jhydrol.2026.135002

Research Groups

• Faculty of Engineering, Ondokuz Mayıs University, Samsun, Turkey
• ZRC SAZU, Karst Research Institute, Postojna, Slovenia
• Institute of Groundwater Management, Technische Universität Dresden, Dresden, Germany

Short Summary

This study developed an innovative hybrid modeling approach, combining the process-based CemaNeige GR6J and data-driven Stacked Autoencoder Deep Neural Networks (SAE-DNN), to robustly predict daily discharge in seasonal snow-influenced karst systems, demonstrating superior performance, especially during extreme flow conditions, compared to standalone models.

Objective

• To develop and test an innovative hybrid modeling approach for robustly predicting daily discharge behavior of karst systems influenced by seasonal snow cover, particularly during extreme flow conditions.

Study Configuration

• Spatial Scale: Unica River catchment, Slovenia, with a total recharge area of approximately 820 square kilometers. The Unica springs are located at 450 meters above sea level, and the main recharge area, the Javorniki Plateau, rises to 1800 meters above sea level.
• Temporal Scale: Daily data covering a 60-year period from 1 January 1962 to 31 December 2021. Analysis also included sub-periods of 10, 20, 30, 40, and 50 years.

Methodology and Data

• Models used:
  • Process-based: G´enie Rural a 6 parametres Journalier (GR6J) coupled with the CemaNeige snow routine (CemaNeige GR6J).
  • Data-driven: Stacked Autoencoder Deep Neural Networks (SAE-DNN).
  • Hybrid: CemaNeige GR6J-SAE-DNN, where SAE-DNN utilizes hydrologically preprocessed outputs (actual evapotranspiration, routing and exponential store components, direct flow) from CemaNeige GR6J as input.
  • Optimization: Differential Evolution (DE) algorithm for model calibration and hyperparameter optimization.
  • Explainable AI: SHAPtree algorithm for feature importance analysis of model parameters and input variables.
• Data sources:
  • Daily precipitation, temperature, potential evapotranspiration, and discharge observation data.
  • Areal precipitation calculated using the Thiessen polygon method from Postojna, Cerknica, and Nova vas Bloke meteorological stations.
  • Potential evapotranspiration, snow accumulation, and melt calculated using temperature data from Postojna and the Oudin formula.
  • Daily discharge data from the Unica Hasberg gauging station.
  • Data publicly available from the Slovenia Environment Agency (ARSO).

Main Results

• The hybrid CemaNeige GR6J-SAE-DNN model significantly outperformed both stand-alone CemaNeige GR6J and SAE-DNN models across all modeling periods.
• The hybrid model improved Nash-Sutcliffe Efficiency (NSE) by 16% (calibration) and 21% (testing) compared to CemaNeige GR6J, and by 174% (calibration/training) and 136% (testing) compared to SAE-DNN.
• The hybrid model demonstrated superior performance in simulating extreme flow conditions (both low and high flows), with its recession constants (average 10.5 days) being more consistent with observed discharge (average 13.5 days) than stand-alone models.
• Model performance varied with sub-periods, showing peak performance for the most recent 10-year period (2012–2021), with NSE and Kling-Gupta Efficiency (KGE) values exceeding 0.90, attributed to the model's ability to adapt to changing climate conditions (e.g., increased temperature, decreased solid precipitation).
• SHAP analysis revealed that the routing store component (QR) was the most impactful input variable for the SAE-DNN within the hybrid structure, while intercatchment exchange parameters (X2 and X5) were most important for the stand-alone CemaNeige GR6J.

Contributions

• Development and successful application of an innovative hybrid modeling approach (CemaNeige GR6J-SAE-DNN) for robust daily discharge prediction in complex, snow-influenced karst systems.
• Demonstration of significant performance enhancement over stand-alone process-based and data-driven models, particularly for extreme flow conditions.
• Utilization of an extensive 60-year historical dataset to evaluate model robustness under changing climate conditions and input variability across multiple time periods.
• Application of SHAP analysis to provide interpretability for both process-based model parameters and data-driven model input variables within the hybrid structure.
• Highlighting the benefit of hydrologically preprocessing input data for deep learning models in complex hydrological systems, offering a practical solution for water resource management in similar karst systems.

Funding

• Scientific and Technological Research Council of Türkiye (TÜB˙ITAK) within the framework of the “TÜB˙ITAK 2219-International Postdoctoral Research Fellowship Program for Turkish Citizens”
• Slovenian Research and Innovation Agency (ARIS) through grants from the Slovenian National Research Programme Karst Research, No. P6-0119

Citation

@article{Sezen2026Robust,
  author = {Sezen, C. and Ravbar, N. and Hartmann, Andreas and Chen, Kai},
  title = {Robust discharge prediction of seasonal snow-influenced karst systems through hybridization of process-based and data-driven models},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135002},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135002}
}