Gohouede et al. (2026) A dimensionally consistent reformulation of the HyMoLAP rainfall–runoff model: interpretability and robustness across CAMELS datasets

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-10-12
• Authors: Lionel Cédric Gohouede, Sianou Ezéckiel Houénafa, Eliézer Iboukoun Biao, Olatunji Olugoke Johnson, Cenk Sezen
• DOI: 10.1016/j.jhydrol.2026.136604

Research Groups

• Department of Mathematics, Pan African University Institute for Basic Sciences, Technology and Innovation, Nairobi, Kenya
• National School of Mathematical Engineering and Modelling, National University of Sciences, Technology, Engineering and Mathematics, Abomey, Benin
• Department of Mathematics, University of Manchester, Manchester, United Kingdom
• Faculty of Engineering, Ondokuz Mayis University, 55139 Samsun, Turkey

Short Summary

This study reformulates the HyMoLAP rainfall–runoff model into d-HyMoLAP by introducing scaling parameters to ensure dimensional consistency. Evaluated across 1172 CAMELS catchments, d-HyMoLAP demonstrates significantly improved robustness and performance over the original HyMoLAP, showing potential to match or exceed the HBV model under specific hydroclimatic conditions.

Objective

• To reformulate the Hydrological Model based on the Least Action Principle (HyMoLAP) by introducing scaling parameters to ensure dimensional consistency, creating d-HyMoLAP.
• To evaluate the performance and robustness of d-HyMoLAP across large-sample CAMELS datasets (CAMELS-FR and CAMELS-GB).
• To investigate the interpretability of d-HyMoLAP performance using machine learning techniques and assess parameter sensitivity and predictive uncertainty.

Study Configuration

• Spatial Scale: Catchment scale, evaluated across 1172 catchments (549 CAMELS-FR and 623 CAMELS-GB).
• Temporal Scale: Utilizes historical rainfall–runoff records for calibration and validation across the selected catchments; stable predictive performance can be achieved with relatively short calibration records.

Methodology and Data

• Models used:
  • Rainfall–runoff models: d-HyMoLAP (reformulated), HyMoLAP (original), HBV (for comparative analysis).
  • Machine learning and sensitivity analysis: Random Forest, XGBoost, SHAP (SHapley Additive exPlanations) analysis, Morris sensitivity analysis, GLUE (Generalized Likelihood Uncertainty Estimation).
• Data sources:
  • CAMELS-FR (Catchment Attributes and MEteorology for Large-sample Studies - France) dataset, comprising 549 catchments.
  • CAMELS-GB (Catchment Attributes and MEteorology for Large-sample Studies - Great Britain) dataset, comprising 623 catchments.

Main Results

• d-HyMoLAP substantially improved performance compared to the original HyMoLAP, increasing the mean validation Nash-Sutcliffe Efficiency (NSE) from 0.165 to 0.649 on CAMELS-FR and from -2.570 to 0.650 on CAMELS-GB.
• The reformulated model significantly reduced catastrophic failures relative to the original HyMoLAP.
• While HBV generally showed stronger performance, d-HyMoLAP can match or exceed HBV in a subset of catchments under specific hydroclimatic conditions.
• In CAMELS-FR, d-HyMoLAP is more likely to perform comparably to or better than HBV in catchments characterized by low runoff ratio, low high-flow frequency, and a gentle flow-duration-curve slope.
• The highest performance of d-HyMoLAP is associated with humid catchments exhibiting stable hydroclimatic conditions and limited hydrological extremes.
• Analysis indicated that stable predictive performance can be achieved using relatively short calibration records.
• Parameter sensitivity analysis identified the newly introduced scaling parameters in d-HyMoLAP as the dominant controls on model performance.
• GLUE analysis for selected catchments showed evidence consistent with practical equifinality and relatively constrained prediction intervals, though coverage was below the nominal 90%.

Contributions

• Presents the first dimensionally consistent reformulation of the HyMoLAP model, named d-HyMoLAP, enhancing its physical interpretability and robustness.
• Demonstrates a substantial improvement in model performance and reduction in catastrophic failures of d-HyMoLAP over the original HyMoLAP across large-sample datasets (1172 CAMELS catchments).
• Identifies specific hydroclimatic and catchment attributes where d-HyMoLAP performs comparably to or better than the established HBV model, providing valuable insights for model application.
• Utilizes advanced machine learning and sensitivity analysis techniques (SHAP, Morris) to provide interpretability into d-HyMoLAP's performance and parameter sensitivity.
• Supports the potential of d-HyMoLAP for large-sample rainfall–runoff modelling and improved water resource management.

Funding

[No specific funding projects, programs, or reference codes were mentioned in the provided paper text.]

Citation

@article{Gohouede2026dimensionally,
  author = {Gohouede, Lionel Cédric and Houénafa, Sianou Ezéckiel and Biao, Eliézer Iboukoun and Johnson, Olatunji Olugoke and Sezen, Cenk},
  title = {A dimensionally consistent reformulation of the HyMoLAP rainfall–runoff model: interpretability and robustness across CAMELS datasets},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136604},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136604}
}