What Role Does Hydrological Science Play in the Age of Machine Learning?

Summary

This commentary, derived from a keynote at Google's 2020 Flood Forecasting Meets Machine Learning Workshop, argues that deep learning methods applied to rainfall-runoff simulation reveal substantially more exploitable information in hydrological datasets than conventional hydrological science has historically translated into theory or models. The authors call for the hydrology community to move beyond subjective, evidence-weak adherence to traditional 'process understanding' and instead develop quantitative frameworks for identifying where process knowledge adds value within machine learning-dominated modelling disciplines.

UK applicability

UK hydrological science and flood forecasting practice could benefit from reassessing reliance on traditional process-based models in favour of hybrid or machine learning approaches, particularly as climate variability increases. The findings are globally applicable to flood risk assessment and water resource management but require integration with UK-specific rainfall and runoff datasets.

Key measures

Comparative predictive performance of deep learning versus conventional hydrological models; information content in large-scale hydrological datasets

Outcomes reported

The study examined how deep learning approaches to rainfall-runoff simulation reveal substantially more information in large-scale hydrological datasets than conventional hydrological models have historically captured. It assessed the tension between machine learning predictive capability and traditional process-based hydrological understanding.

Topic tags