Model calibration and improved process understanding

Beven et al. (2022) present a novel invalidation approach to calibrate an ensemble of Dynamic Topmodel parameter sets in a study examining the potential for hillslope storage bunds to mitigate the effects of downstream flooding in the 209 km2 River Kent catchment in UK. The model invalidation approach is based on a GLUE-like methodology, where the acceptability thresholds in a first goodness-of-fit step is defined to reflect the uncertainty associated with input and discharge data. 118 realizations out of 100,000 survived evaluations of hydrograph peaks in three years with major floods. While most model calibrations are confined to such goodness-of-fit measures of how well models perform in discharge simulations, Beven et al. (2022) introduced an additional evaluation step, where only the simulations with at least 10% of the area producing overland flow during the largest storm were accepted. This fitness-for-purpose measure reduced the acceptable realizations to 67. Altogether, the 67 surviving realizations are not necessarily those that give the highest Nash-Sutcliffe efficiency values, but those that are considered most suitable for assessing the impact of certain flood mitigation measures in the catchment.
De Lavenne et al. (2022) use the HYPE model for 111 catchments spread across the USA to evaluate the effect of calibration against both discharge and sediment data instead of only discharge data and to evaluate five hypotheses for overland flow process descriptions. The results confirm previous findings that inclusion of a second data set (in this case sediment) in a multi-objective calibration approach generally lead to significantly improved simulations for sediment concentrations with only a slightly reduced performance for discharge. The five overland flow modelling hypotheses consist of the existing formulation using three parameters and four new formulations using one or two parameters. The results show that the performance for discharge simulations is not improved by the new hypotheses, while the performances for sediment concentrations are improved. In addition, equifinality is reduced by the new hypothesis due to a lower number of model parameters.
La Folette et al. (2022) study streamflow simulation in the 16.9 km2 Elder Creek catchment in Northern California, where the geology is characterized by fractured bedrock overlain by a, typically thin (0.5 m), soil layer. This is the first study where unsaturated weathered bedrock water storage is explicitly incorporated in a catchment model and used as a calibration target. They calibrate a lumped rainfall-runoff model against three observations targets: i) only streamflow data; ii) only rock moisture data; and iii) both streamflow and rock moisture data. The calibration is performed by evaluating 10,000 parameter sets using the concept of pareto optimality. The results show that the model calibrated against both streamflow and rock moisture data is capable of accurately simulating the dynamics in rock moisture and streamflow, while a calibration against streamflow data alone may result in relatively poor simulation of rock moisture dynamics and a calibration against rock moisture alone may result in relatively poor simulation of streamflow dynamics. Furthermore, the results show that the calibrated parameter values appear more physically realistic when calibrating against both streamflow and rock moisture data. The study concludes that incorporation of rock moisture data can lead to a more robust model, that without sacrificing the accuracy of streamflow simulations can provide increased accuracy of some model results and decreased parameter uncertainty.