This paper is derived from a keynote talk given at the Google's 2020 Flood Forecasting Meets Machine Learning Workshop. Recent experiments applying deep learning to rainfall‐runoff simulation indicate that there is significantly more information in large‐scale hydrological data sets than hydrologists have been able to translate into theory or models. While there is a growing interest in machine learning in the hydrological sciences community, in many ways, our community still holds deeply subjective and nonevidence‐based preferences for models based on a certain type of “process understanding” that has historically not translated into accurate theory, models, or predictions. This commentary is a call to action for the hydrology community to focus on developing a quantitative understanding of where and when hydrological process understanding is valuable in a modeling discipline increasingly dominated by machine learning. We offer some potential perspectives and preliminary examples about how this might be accomplished.

The most accurate rainfall–runoff predictions are currently based on deep learning. There is a concern among hydrologists that the predictive accuracy of data-driven models based on deep learning may not be reliable in extrapolation or for predicting extreme events. This study tests that hypothesis using long short-term memory (LSTM) networks and an LSTM variant that is architecturally constrained to conserve mass. The LSTM network (and the mass-conserving LSTM variant) remained relatively accurate in predicting extreme (high-return-period) events compared with both a conceptual model (the Sacramento Model) and a process-based model (the US National Water Model), even when extreme events were not included in the training period. Adding mass balance constraints to the data-driven model (LSTM) reduced model skill during extreme events.

Ingesting near-real-time observation data is a critical component of many operational hydrological forecasting systems. In this paper, we compare two strategies for ingesting near-real-time streamflow observations into long short-term memory (LSTM) rainfall–runoff models: autoregression (a forward method) and variational data assimilation. Autoregression is both more accurate and more computationally efficient than data assimilation. Autoregression is sensitive to missing data, however an appropriate (and simple) training strategy mitigates this problem. We introduce a data assimilation procedure for recurrent deep learning models that uses backpropagation to make the state updates.

Accurate representation of the turbulent exchange of carbon, water, and heat between the land surface and the atmosphere is critical for modelling global energy, water, and carbon cycles in both future climate projections and weather forecasts. Evaluation of models' ability to do this is performed in a wide range of simulation environments, often without explicit consideration of the degree of observational constraint or uncertainty and typically without quantification of benchmark performance expectations. We describe a Model Intercomparison Project (MIP) that attempts to resolve these shortcomings, comparing the surface turbulent heat flux predictions of around 20 different land models provided with in situ meteorological forcing evaluated with measured surface fluxes using quality-controlled data from 170 eddy-covariance-based flux tower sites. Predictions from seven out-of-sample empirical models are used to quantify the information available to land models in their forcing data and so the potential for land model performance improvement. Sites with unusual behaviour, complicated processes, poor data quality, or uncommon flux magnitude are more difficult to predict for both mechanistic and empirical models, providing a means of fairer assessment of land model performance. When examining observational uncertainty, model performance does not appear to improve in low-turbulence periods or with energy-balance-corrected flux tower data, and indeed some results raise questions about whether the energy balance correction process itself is appropriate. In all cases the results are broadly consistent, with simple out-of-sample empirical models, including linear regression, comfortably outperforming mechanistic land models. In all but two cases, latent heat flux and net ecosystem exchange of CO₂ are better predicted by land models than sensible heat flux, despite it seeming to have fewer physical controlling processes. Land models that are implemented in Earth system models also appear to perform notably better than stand-alone ecosystem (including demographic) models, at least in terms of the fluxes examined here. The approach we outline enables isolation of the locations and conditions under which model developers can know that a land model can improve, allowing information pathways and discrete parameterisations in models to be identified and targeted for future model development.