Advancing distributed hydrological modeling with hybrid machinelearning

Accurately simulating large-scale water dynamics is important

for managing water resources, addressing climate change impacts, and

understanding hydrological variability. Despite advances in hydrological

modeling, simulating water fluxes and states at global or regional

scales remains challenging due to the complexity of distributed

processes and limited understanding of key components. Encoding physical

knowledge in deep neural networks (NNs) for differentiable modeling

offers a promising solution but has yet to be fully realized for

distributed hydrological models, especially for processes such as river

routing.

This study presents a novel differentiable modeling framework that

bridges physical and data-driven approaches for distributed hydrological

modeling. The framework encodes a large-scale hydrological model (i.e.,

HydroPy) as a neural network, incorporates an additional NN to map

spatially distributed parameters from local climate and land attributes,

and employs NN-based modules to represent poorly understood processes.

Multi-source observations are used to constrain the system in an

end-to-end manner, with the Amazon Basin as a case study to demonstrate

the framework’s applicability and effectiveness.

Results show that the developed model improves simulation accuracy by

30-40% compared to the original hydrological model. Replacing the

Penman-Monteith formulation with NN produces more realistic potential

evapotranspiration estimates. SHAP analysis of the NN parameterization

further reveals how climate and land attributes regulate the spatial

variability of key parameters. Overall, by integrating physical realism

with the flexibility of machine learning, this framework addresses

critical limitations of traditional hydrological models. It provides a

scalable, interpretable approach to advance large-scale hydrological

modeling and address pressing water and climate challenges.