Evaluation of the structural uncertainty of hydrological models in the estimation of multiple hydrological processes

Hydrological models are powerful tools, that allows users to create a simplified representation of real-world system and that serve to understand the hydrological processes in a basin, and predict their future behavior, including, for example, the effects of climate change. However, these models are subject to multiple sources of uncertainty, including structural uncertainty, related to the hydrological processes simulated and to the spatial discretization applied (lumped, semi-distributed or distributed models). The effects of this modelling decision could be particularly relevant when the objective is to simulate more than one hydrological process.

The objective of this work is to determine if the use of different model structures (lumped or semi-distributed), and the selection of process to be simulated allows reducing the uncertainty of the estimation of more than one hydrological process. Using Raven, a robust and flexible hydrological modelling framework, that supports a wide variety of modelling options, and sits atop a robust and extendible software architecture, eight model structures have been constructed to simulate the River Colorado en Junta con Palos Basin. This basin located in the central zone of Chile (Lat.-35.25, Lon. 71), has a snow-pluvial regime and an average annual rainfall of 1796 mm for the period 1979-2020.  Additionally, this basin covers an area of 879 km², with a wide elevation range, from 643 m.a.s.l. to 4074 m.a.s.l.

The results have shown some differences at modelling daily streamflow (KGE from 0.68 to 0.72 in the lumped models, and from 0.68 to 0.8 in semi-distributed models). Furthermore, other important changes have been visualized related to the characterization of snow cover and soil moisture in the first layer of soil. The simulated series have been compared to satellite data (products MODIS10A2 for snow cover and NASA-USDA Enhanced SMAP Global Soil Moisture Data for superficial soil moisture).

In the case of the snow cover, the annual duration of snow cover was evaluated, obtaining Pearson's coefficient values between 0.4 and 0.56 for lumped models, while these values reach 0.65 in the case of semi-distributed models. Regarding soil moisture, the changes were more significant when changing the structure of the model (selection and parameterizations of the processes), rather than its spatial discretization, with a range of KGE values from 0.34 to values close to 0.7, strongly influenced by the methods used to evaluate evapotranspiration and infiltration, as well as by the characteristics of the soil.

Overall, this work demonstrates the potential of a flexible hydrologic modeling framework to assess and reduce the structural uncertainty of hydrologic models, taking advantage of the potential of these tools.