 |
| Titel |
Data mining methods for predicting event runoff coefficients in ungauged basins using static and dynamic catchment characteristics |
| VerfasserIn |
Ralf Loritz, Markus Weiler, Simon Seibert |
| Konferenz |
EGU General Assembly 2015
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 17 (2015) |
| Datensatznummer |
250111011
|
| Publikation (Nr.) |
 EGU/EGU2015-11072.pdf |
|
|
|
|
|
| Zusammenfassung |
| Transferring hydrological information into ungauged basin by regionalisation approaches is an ongoing field
of research. Usually regionalisation techniques use physical landscape descriptors to transfer either model
parameters or hydrological characteristics from a catchment to another. A common problem of these
approaches is the high degree of uncertainty associated to their results. One reason is that often solely static
(structural) catchment characteristics such as catchment area, physiographic properties or land use data are
used for regionalisation. However, it is well known that the hydrological response of a 'natural' system is a
complex and a non-linear interaction of its structure, state and forcing. Here it is important to note, that only
structure is a static property. State and forcing are highly dynamic when considering the temporal and spatial
scale of a rainfall-runoff event.
To overcome the limitations associated with 'static' regionalisation techniques we propose a regionalisation
technique for event runoff coefficients combining static and dynamic catchment properties. The approach is
based on the two data mining algorithms 'random forests' and 'quantile regression forests'. The static
catchment characteristics include standard variables such as physiographic properties, land cover and soil
data. The dynamic variables include event based properties of the forcing (i.e. rainfall amount, intensity,...)
and proxies for the initial state of the catchment (i.e. initial soil moisture). Together with the runoff
coefficient these quantities were extracted form hydro-meteorological time series (precipitation, discharge
and soil moisture) using an automated rainfall-runoff event detection technique.
We tested our method using a set of 60 meso-scale catchments (3.1 to 205,6 km2, covering a range of
different geologies and land uses) from Southwest Germany. We randomly separated the catchments in two
groups. The first group (30 donor catchments) was used to train the data mining models . Based on the
resulting relations we then predicted event runoff coefficients for the other half of the catchments (30 test
catchments). With this regression method we are able to predict event runoff coefficient s in the test group
with an overall root mean square error of about 5%. Furthermore our approach indicates that the dynamic
characteristics (event precipitation and initial soil moisture) had a much higher importance for the prediction
of event runoff coefficients than the static properties. In the next step, we applied random forest regressions
to all 60 catchments individually based on the extracted event variables. The relative importance of the
predictor variables of each of these regressions can be interpreted as indicators for the dominating rainfall-
runoff controls within the basins (e.g. to identify initial storage or rainfall intensity controlled conditions).
We conclude that the ensemble regression tree methods provide insights into the 'functioning' of the
individual catchments and that dynamic catchment properties (observed on meaningful spatial and temporal
scales) have a very high potential for prediction hydrological functions in ungauged catchments. |
|
|
|
|
|
|