HR: 0800h
AN: H51B-0358 [Abstracts]
TI: Tapping the Information Content in Rainfall-Runoff Data by Partitioning With Regression
Trees
AU: * Iorgulescu, I
EM: Ion.Iorgulescu@epfl.ch
AF: Institute of Environmental Science and Technology, Ecole Polytechnique
Fédérale de Lausanne, EPFL, ENAC, ISTE,
GECOS, Lausanne, CH-1015
Switzerland
AU: Beven, K J
EM: k.beven@lancaster.ac.uk
AF: Institute of Environmental and Natural Sciences, Lancaster University, Lancaster University, Lancaster,
LA1 4YQ
United Kingdom
AB:
The current approach in rainfall-runoff modeling that assumes global hydrologic and error models is successful when the
latter are not to far from the 'true' ones. As we argue elsewhere, this will likely not be the case for applications on real
data. Our main assumption is that there is more information than previously thought in long, good quality, rainfall-runoff
records. We suggest that there is a potential tradeoff between the quantity of data and the 'strength' of the adopted
modeling hypotheses. We will present and discuss a new partitioning method that is based on the "nonparametric direct mapping
of rainfall-runoff relationships" approach we recently introduced. This method is based on the recursive binary partitioning
of input space using regression trees, a non-linear, non-parametric, identification algorithm. We will show that the latter
improves the consistency of some otherwise subjective choices, such as the objective functions or 'behavioural' thresholds
used for model evaluation. A first application of the proposed methodology on real data will be presented.
DE: 1846 Model calibration (3333)
DE: 1847 Modeling
DE: 1873 Uncertainty assessment (3275)
DE: 1879 Watershed
SC: Hydrology [H]
MN: Fall Meeting 2005