Developing an event-based distributed hydrological model through competing hypotheses and meta-metrics of performance

Event-scale hydrological modeling applications entail fine temporal discretization, enhanced model components, and carefully refined initial and boundary conditions. However, realistic modeling requires justifying assumptions that influence model complexity and the dominant processes represented for a specific catchment. This process is particularly challenging for distributed hydrological models, which, compared to lumped models, incorporate additional assumptions to account for spatial variability in hydrological processes.

This study demonstrates a modeling approach that uses controlled comparisons and meta-metrics of performance to develop a distributed model for a semi-arid catchment in Southern Italy. From hydrological signatures we hypothesize that in this catchment both Hortonian (infiltration excess) and Dunnian (saturation excess) runoff mechanisms can concurrently appear in hydrograph responses during rainfall events. Our objective is to disentangle these mechanisms and design a model capable of distinguishing between them. We therefore developed four perceptual model architectures representing different runoff generation hypotheses, informed by hydrological signatures, and tested them within a nested catchment framework.

A multi-stage operational test involving the calibration of a meta-objective function and spatial transferability validation was conducted to provide a robust and unequivocal ranking of the best-performing models, exposing unsolved structural problems of competing hypotheses. Assessing the consequences of simulated high flows by replacing 2D Shallow Water equations to a simplified routing scheme reinforces the idea of replacing popular metrics with meta-metrics.

Posterior diagnostics confirmed that the most realistic model structure, as indicated by internal consistency in simulated processes, aligned with the highest meta-metrics performance. Hydrographs comparison and hypotheses falsification further revealed that the dominant runoff mechanisms during consecutive storm events could be clearly disentangled, with Hortonian and Dunnian processes alternating depending on rainfall intensity and soil wetness.

By integrating multiple working hypotheses with enhanced operational testing, our proposed model development approach shows that even with limited observational data, such as sole streamflow measurements within a nested catchment setup, it is possible to identify runoff generation processes in event-scale hydrological applications.