Experiences with transferability and scaling of hydrologic model parameters

Hydrologic models, regardless of their level of conceptualization of the underlying physical mechanisms, inevitably require some parameter estimation or calibration. A trend toward increased physical realism in hydrologic models has yet to be shown to reduce, or ideally eliminate, the need for model calibration. Furthermore, with the evolution of macroscale hydrological models, which may be tied to specific grid resolutions, questions arise as to the transferability of model parameters among different resolutions. Results are shown based on experience with two hydrologic models appropriate to somewhat different spatial scales. The Distributed Hydrology-Soil-Vegetation Model (DHSVM) is a topographically explicit spatially distributed hydrologic model that represents runoff generation (primarily via the saturation excess mechanism) and other surface hydrologic processes over grid resolutions typically ranging from about 30 to 150 m. It has been applied to a number of catchments within the Puget Sound drainage basin, and is used to produce experimental flood forecasts at about 60 locations within the basin. As part of the model implementation effort, manual parameter estimation was performed for the Snoqualmie River basin (about 2000 km2), and model parameters were then transferred to other catchments within the region. Evaluation of initial simulations and forecasts based on the transferred parameters are shown, in comparison with simulations made using locally estimated parameters. The results are generally quite encouraging. In the arena of macroscale hydrology, where models like the Variable Infiltration Capacity (VIC) model are often designed for coupling with atmospheric models, the transferability problem arises due to frequent changes in spatial resolution of the atmospheric model, typically brought about due to advances in computing power. Simulations were run with the VIC model run off-line over the Columbia and Arkansas-Red River basins, at spatial resolutions from 1/8 to 2 degrees latitude by longitude. Analysis of the simulations showed that predicted mean annual streamflow was as much as 18 and 12 percent lower for the Arkansas-Red and Columbia basins, respectively, at the lowest as compared with the highest spatial resolution. When subgrid spatial variability, and variations of precipitation with elevation were parameterized, the maximum sensitivity of annual average runoff was reduced from 18 to 14 percent in the Arkansas-Red, and from 12 to 4 percent in the Columbia.