RAL SEMINAR SERIES
THURSDAY, MAY 11, 2017 | 2PM-3PM | FL2 - 1001
NATIONAL CENTER FOR ATMOSPHERIC RESEARCH, BOULDER
A Bayesian Approach to Address the Structural Uncertainty of Hydrological Models in Ungauged Catchments
CRISTINA PRIETO | ENVIRONMENTAL HYDRAULICS INSTITUTE “IHCANTABRIA”. UNIVERSITY OF CANTABRIA, SANTANDER, SPAIN
Reliable predictions of available water resources are essential for water management and its related economic activities. These predictions typically come from rainfall runoff models, yet in the vast majority of the catchments across the world, no streamflow measurements are available. Producing reliable predictions in ungauged catchments remains a major research and applied challenge for the hydrological community.
This talk will focus on two aspects of predictions in ungauged catchments: 1) the assumption that the pre-selected model and regionalization technique are a suitable and adequate representation of the behavior in the ungauged catchment, and 2) the identification of dominant hydrological mechanisms to be represented in hydrological models.
First, hydrological indices (also named hydrological signatures) are selected using a principal components analysis. Second, the most significant principal components are regionalized using random forests regression techniques. Third, these regionalized principal components are assimilated into hydrological models using a Bayesian approach. Fourth, two metrics are proposed to assess model and regionalization suitability and adequacy. Suitability measures whether the model and regionalization are able to reproduce catchment behavior with a high probability, while adequacy measures the relative gain of knowledge of including a model or regionalization procedure. And fifth, a new procedure is introduced, which enables the identification of dominant hydrological mechanisms from regionalized information and an ensemble of models using a Bayesian approach.
The methodological developments have been applied to basins in northern Spain with diverse hydroclimatological regimes. The results indicate that the prediction quality is sensitive to the model error, to the quality of the regionalized information, and to the available information content.