Zhou et al. (2026) HydroTreeNet: differentiable hydrological process trees for physically consistent rainfall–runoff modeling

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-10-07
• Authors: Xuerui Zhou, Baowei Yan, Jianbo Chang, Shixiong Du, Dongxu Yang, Mingbo Sun
• DOI: 10.1016/j.jhydrol.2026.136509

Research Groups

• School of Civil and Hydraulic Engineering, Huazhong University of Science and Technology, Wuhan, China
• Hubei Key Laboratory of Digital River Basin Science and Technology, Huazhong University of Science and Technology, Wuhan, China

Short Summary

This paper introduces HydroTreeNet, a novel process-constrained differentiable rainfall-runoff framework that integrates spatially heterogeneous runoff generation, constrained flux partitioning, and explicit river-network routing. The model demonstrated competitive, albeit basin-dependent, predictive performance with high mass consistency across various catchment types, highlighting the importance of snow processes and explicit routing.

Objective

• To develop and evaluate HydroTreeNet, a process-constrained differentiable rainfall-runoff framework that explicitly integrates node-scale snow and liquid-water storages, generalized complementary allocation at hydrological interfaces, and explicit routing over a mapped upstream-downstream river network, while maintaining mass consistency.

Study Configuration

• Spatial Scale: Nine diagnostically selected CAMELS-SPAT catchments.
• Temporal Scale: Continuous simulation over multiple calibration and validation periods (implied by repeated calibrations and validation-tuned benchmarks).

Methodology and Data

• Models used: HydroTreeNet (proposed framework), neural and conceptual benchmarks (for comparison).
• Data sources: CAMELS-SPAT catchments (implying meteorological forcing and hydrological observations).

Main Results

• Predictive performance was strongly basin dependent, with no single model dominating all catchments.
• HydroTreeNet was competitive in selected Mediterranean, humid, rain–snow, and large lowland-network settings.
• The model exhibited a pronounced structural failure in an intermittent pulse-flow basin.
• Removing snow processes reduced Nash-Sutcliffe Efficiency (NSE) by 0.32–0.41 in two snow-influenced basins.
• Removing explicit network routing reduced NSE by up to 0.20 in several selected basins.
• Whole-system water-balance residuals remained below 2 × 10⁻⁷ relative to cumulative input, demonstrating high mass consistency.
• Repeated-seed differences in internal states revealed persistent equifinality despite similar outlet skill.

Contributions

• Introduces HydroTreeNet, a novel process-constrained differentiable rainfall–runoff framework that jointly represents spatially heterogeneous runoff generation, constrained flux partitioning, and explicit river-network propagation.
• Achieves high mass consistency with whole-system water-balance residuals below 2 × 10⁻⁷.
• Provides an auditable framework for separating predictive behavior, structural hypotheses, and numerical conservation within differentiable rainfall–runoff modeling.
• Quantifies the impact of specific process components (snow, explicit network routing) on predictive performance.

Funding

• Not specified in the provided text.

Citation

@article{Zhou2026HydroTreeNet,
  author = {Zhou, Xuerui and Yan, Baowei and Chang, Jianbo and Du, Shixiong and Yang, Dongxu and Sun, Mingbo},
  title = {HydroTreeNet: differentiable hydrological process trees for physically consistent rainfall–runoff modeling},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136509},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136509}
}