eWaterCycle in Practice

Abstract

This dissertation addresses multiple challenges in hydrological modeling, focusing on reproducibility, large-sample catchment diagnostics, and the role of observational uncertainty in interpreting model performance. Through a series of applied studies, it proposes practical tools and methods to improve the transparency, comparability, and interpretability of hydrological models, without overstating what current data or models can reliably deliver. Throughout, the thesis argues that apparent limitations in model evaluation, including the persistence of equifinality, can be reduced by expanding the observational basis against which models are tested, rather than accepting them as inherent properties of the system.

Central to this work are the development and application of the eWaterCycle platform: a modular, community-driven system that promotes reproducible modeling\\ through standardized interfaces and containerized environments. By building on existing models through the Basic Model Interface (BMI) and deploying experiments at scale, eWaterCycle enables model intercomparison and refinement with fewer technical barriers. Although it does not eliminate all integration complexity, the platform provides a practical foundation for reproducible, FAIR-compliant workflows and has already enabled new forms of collaborative experimentation.

Using this platform, the thesis evaluates the effects of spatial resolution on hydrological model performance. Results from 454 CAMELS-US catchments show that a finer spatial resolution (down to 200~m) does not universally improve streamflow predictions at catchment outlets. Although resolution matters in some contexts, particularly where topography or land cover is complex, the overall gains are limited and often fall within uncertainty bounds. These findings suggest that investing in higher spatial resolution without a corresponding investment in observational density is a questionable allocation of research effort. ``Hyper-resolution'' modeling, though technically feasible, must be approached with clear use cases and a realistic assessment of whether the available data can actually support it.

Expanding the scope of model diagnostics, the research applies machine learning (specifically, a random forest) to link catchment attributes to model performance across 617 CAMELS-GB catchments. The analysis shows that model suitability often varies with catchment scale and physical characteristics, with different models (wflow\_sbm and PCR-GLOBWB) excelling under different conditions. This underscores the ``reality of locality'' in hydrology: every catchment presents a unique combination of dominant processes, making the pursuit of a universal model impractical. Although the results expose structural model limitations, they also demonstrate the potential of machine learning as a diagnostic tool, an application that deserves greater emphasis alongside its more common use in prediction.

A key contribution of the dissertation is the integration of uncertainty in discharge observations into model evaluation. Using quantified uncertainty estimates from 299 CAMELS-GB catchments, the analysis shows that a significant share of perceived performance improvements, whether achieved through parameter tuning or arising from structural differences, may fall within the uncertainty bounds of the observations. Notably, no clear relationship was found between model complexity and model performance, challenging the assumption that more complex models are inherently more capable. Together, these findings call into question how model results are interpreted and underline the need to report and account for uncertainty more systematically.

The concluding chapters synthesize these results, arguing for a more critical yet constructive modeling practice: one that embraces reproducibility and large-sample evaluation while recognizing the limits of model resolution, structural complexity, and observational quality. They make the case that advancing hydrology depends as much on investing in new and better observations---from rehabilitating existing gauging networks to deploying novel sensing technologies and satellite products---as it does on improving models. By embedding uncertainty diagnostics into the evaluation process and facilitating open, repeatable workflows, this work contributes to a more robust and realistic foundation for model improvement and application.

Looking ahead, the thesis outlines clear directions for future research: expanding eWaterCycle to include more models and domains; exploiting underutilized satellite missions and opportunistic data sources; and bridging the cultural divide between modelers and field scientists so that data collection and model experimentation become two sides of the same scientific process. Although challenges remain, particularly in the community-wide adoption of reproducible practices, the tools and insights developed here offer a tangible step forward.

In summary, this dissertation contributes to hydrological science by combining reproducible modeling tools with diagnostic depth. It promotes a more transparent and critical modeling practice, one that acknowledges uncertainty, challenges complacency about data scarcity, and encourages the responsible use of models. Rather than seeking breakthroughs, the work focuses on enabling more reproducible, interpretable, and broadly applicable hydrological modeling, thereby helping build a stronger foundation for both scientific progress and practical decision-making. The path forward is not to accept the limitations we have inherited, but to systematically reduce them.