Home > Research > Publications & Outputs > Regionalization as a learning process

Associated organisational unit

View graph of relations

Regionalization as a learning process

Research output: Contribution to Journal/MagazineJournal articlepeer-review

Article numberARTN W11419
<mark>Journal publication date</mark>17/11/2009
<mark>Journal</mark>Water Resources Research
Issue number11
Number of pages13
Publication StatusPublished
<mark>Original language</mark>English


This paper deals with the uncertainty involved in geographical migration of hydrological model structures, commonly known as regionalization. Regionalization relies on the hypothesis that calibrated parameter sets from a donor catchment can be useful to predict discharge of an ungauged catchment. However, since every catchment is unique, model parameters need to be adapted for differences between a calibration and a prediction catchment, either by transformation or further selection. This process is inherently uncertain. Model parameters, and therefore the required changes, do not exactly represent quantities that we can measure or calculate. This paper outlines an approach to learn about how model parameters should be transformed between a gauged and an ungauged catchment. The approach consists of an iterative process, in which a model structure is applied successively to gauged catchments. After each step, parameter behavior is evaluated as a function of catchment properties and intercatchment similarities. The method is illustrated with an application of a customized version of TOPMODEL to a set of catchments in the Ecuadorian Andes. First, parameter sets are generated for a donor catchment. This model ensemble is then used to predict the discharge of the other catchments, after applying a stochastic parameter transformation to account for the uncertainty in the model migration. The parameter transformation is then evaluated and improved before further application. The case study shows that accurate predictions can be made for predicted basins. At the same time, knowledge is gained about model behavior and potential model limitations.