Can Deep Learning Revolutionize Hydrology?

Process-based models, such as land surface and hydrologic models (LSMs/HMs), have been foundational to hydrological research for decades. These models are grounded in the principles of mass, energy, and momentum conservation, providing critical insights into the terrestrial water cycle and forming an essential component of Earth System Models. Despite their importance, process-based models face significant limitations, primarily due to parametric and structural uncertainties that hinder their transferability across scales and locations, ultimately reducing their predictive accuracy.

In contrast, machine learning (ML) models learn directly from data, offering potential advantages for capturing highly nonlinear and complex processes, especially when large datasets are available. However, ML models also have notable drawbacks, including a lack of interpretability (often regarded as "black-box" models, despite efforts to develop more explainable or "physically aware" variants), dependence on data quality and availability, and challenges in generalizing under climate or environmental change conditions.

Given the rapid adoption of ML techniques in recent hydrological literature, a key question arises: Can deep learning replace traditional hydrological models due to its speed and accuracy, or is this shift merely a transient trend?

In this presentation, I will argue that before addressing this question, it is essential to establish two key prerequisites: (1) the purpose of the modeling effort, and (2) the appropriate protocols and metrics [1,2] for evaluating model efficiency. To formalize this discussion, I will propose a set of postulations for each modeling paradigm. Drawing on several examples, I will suggest that the most promising future lies in hybrid modeling frameworks, where the empirical aspects of LSMs/HMs (e.g., pedo-transfer function derivation) could be augmented by ML techniques [3,4], while maintaining the core physical processes [5]. ML could also serve as a valuable tool for estimating human-made impacts [6] on the hydrological system, where first-principles models are often lacking.

References:

[1] Rakovec et al. https://doi.org/10.1002/2016WR019430    
[2] Samaniego et al. https://doi.org/10.5194/hess-21-4323-2017
[3] Feigl et al. https://doi.org/10.1029/2022WR031966
[4] Li et al. https://doi.org/10.1029/ 2023WR035543
[5] Kholis et al. https://doi.org/10.22541/essoar.173532490.04454195/v1
[6] Shrestha et al.  https://doi.org/10.1029/ 2023WR035433