Degenne et al. (2026) Testing the temporal and spatial transferability of a water balance model using a parameter learning approach
Identification
- Journal: Hydrology and earth system sciences
- Year: 2026
- Date: 2026-11-16
- Authors: Antoine Degenne, François Bourgin, Charles Perrin, Vazken Andréassian
- DOI: 10.5194/hess-30-7061-2026
Research Groups
- HYCAR Research Unit, INRAE, Université Paris-Saclay, Antony, France
Short Summary
This study investigates the temporal and spatial transferability of a differentiable hybrid annual water balance model across 3044 catchments from eight countries. It finds that the model excels in temporal transferability with physically coherent parameters but struggles with spatial transferability due to difficulties in regionalizing the long-term mean streamflow (MQ).
Objective
- To investigate whether a differentiable hybrid approach can identify robust and physically coherent parameter sets for annual hydrological modeling, specifically assessing its temporal and spatial transferability using an end-to-end training framework.
Study Configuration
- Spatial Scale: 3044 catchments from eight countries (United States, Australia, Switzerland, Germany, Denmark, France, Great Britain, Sweden), representing approximately 90,000 station-years of data.
- Temporal Scale: Annual time step, with cross-validation periods spanning 1989–1998, 1999–2008, and 2009–2021. Each catchment included at least six complete hydrological years per period.
Methodology and Data
- Models used:
- A differentiable hybrid framework combining a physical water balance model with a dense neural network.
- The physical model consists of an annual anomaly model of climate elasticities (linear regression with three elasticity coefficients: eQ/P, eQ/E, eQ/λ) coupled with the Turc–Mezentsev formulation for the long-term streamflow mean (MQTM), adjusted by a multiplicative coefficient (eMQ). The full model is Qn = eQ/P ΔPn + eQ/E ΔE0n + eQ/λ Δλn + eMQ MQTM.
- A dense neural network (8 hidden layers of 192 units each, approximately 2.7 × 10^5 trainable weights) maps 27 catchment descriptors to the four model parameters (eQ/P, eQ/E, eQ/λ, eMQ) in an end-to-end fashion.
- Optimization: AdamW algorithm, minimizing the Mean Squared Error (MSE) between simulated and observed annual streamflow (in mm yr⁻¹).
- Benchmark: Local, catchment-by-catchment linear regressions using ordinary least squares.
- Data sources:
- Large-sample dataset of 3044 catchments from various CAMELS (Catchment Attributes and Meteorology for Large-sample Studies) datasets.
- Climatic data (precipitation, potential evaporation) from original datasets, with potential evaporation recomputed using the Oudin et al. (2005) formula.
- 27 catchment descriptors (9 climatic, 7 land cover/human influence, 11 subsurface/lithological/geometric) derived from global data products (e.g., GLC2000, HID v1.0, GLHYMPS, GLIM).
Main Results
- Temporal Transferability: The hybrid model, trained globally, significantly outperforms local, catchment-by-catchment calibration in temporal cross-validation, achieving higher Nash-Sutcliffe efficiency (NSE) values.
- Parameter Coherence: The hybrid model implicitly learns physically plausible elasticity coefficients (e.g., eQ/P in [0, 1], eQ/E and eQ/λ in [-1, 0]) without explicit constraints, unlike local regressions which often set coefficients to zero due to statistical insignificance.
- Spatial Transferability Limitations: Model performance drops substantially in spatial and spatiotemporal cross-validation compared to temporal cross-validation, indicating a primary challenge in spatial rather than temporal transfer.
- MQ Regionalization Bottleneck: The elasticity component of the model (streamflow anomalies) regionalizes well across space. However, the main loss of performance in spatial transferability is attributed to the model's difficulty in accurately predicting the long-term mean streamflow (MQ) from catchment descriptors, with MQ prediction errors increasing markedly under spatial cross-validation.
- Role of Descriptors: In temporal cross-validation, descriptors primarily act as catchment identifiers. In contrast, for spatial cross-validation (regionalization), the hybrid model critically relies on the physical content of descriptors; replacing them with random vectors leads to a significant performance degradation and reduced parameter variability.
Contributions
- First application of a differentiable hybrid approach to a parsimonious annual-scale water balance model, demonstrating its potential for learning robust and interpretable anomaly parameters.
- Provides evidence that global, end-to-end training in a differentiable hybrid framework improves temporal transferability and yields physically coherent parameter sets without explicit constraints.
- Identifies the regionalization of the long-term mean streamflow (MQ) as the primary bottleneck for achieving spatially transferable annual streamflow predictions with this type of model.
- Clarifies the distinct roles of catchment descriptors in temporal (as identifiers) versus spatial (as carriers of physical information) transferability for hybrid models, contrasting with findings for purely data-driven models.
Funding
- Agence Nationale de la Recherche (project DRHYM ANR-22-CE56-0007).
- HPC resources from GENCI–IDRIS (grant nos. 2024-AD010115988 and 2025-AD010115988R1).
Citation
@article{Degenne2026Testing,
author = {Degenne, Antoine and Bourgin, François and Perrin, Charles and Andréassian, Vazken},
title = {Testing the temporal and spatial transferability of a water balance model using a parameter learning approach},
journal = {Hydrology and earth system sciences},
year = {2026},
doi = {10.5194/hess-30-7061-2026},
url = {https://doi.org/10.5194/hess-30-7061-2026}
}
Original Source: https://doi.org/10.5194/hess-30-7061-2026