 |
| Titel |
Analyzing and modeling complex weather radar data with data-driven approaches |
| VerfasserIn |
Reinhard Teschl, Franz Teschl, Walter L. Randeu |
| Konferenz |
EGU General Assembly 2013
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 15 (2013) |
| Datensatznummer |
250082065
|
|
|
|
|
|
| Zusammenfassung |
| In the field of radar hydrology the utilization of data-driven models seems promising because
the data volume produced by weather radar networks is considerably large. Reams of
gigabytes of data are stored in the archives. However, these complex datasets are not easy to
investigate. Data-driven approaches aim to extract and model patterns and regularities that are
hidden in the datasets.
This study presents data-driven models for three aspects of radar hydrology: data
analysis, rainfall-runoff prediction, and radar rainfall estimation. The Principle Component
Analysis (PCA) has been used to capture the essence in weather radar measurements and to
provide methods for describing patterns in the spatial radar data. For this analysis, volumes
that are scanned concurrently by two radar stations of the Austrian weather radar network
were used for plausibility reasons.
Artificial Neural Networks (ANNs) were applied to predict the runoff of a small Alpine
catchment. Several input configurations and network architectures were investigated. The
models were trained on various lead times and the ANNs consistently perform better than
simpler approaches like Model Trees (MTs) applied on the same dataset. When forecasting
three time steps ahead, the ANN model reaches an efficiency coefficient of 97.4 % compared
to 90.9 % of the MT.
Data-driven models were also used to improve weather radar estimates of rainfall. By
means of ANNs the radar reflectivity Z above a rain gauge was mapped to the rain rate R on
the ground. The so modeled relationship was tested on a different location. The
deviations could be decreased and the correlation coefficient increased compared to
applying the standard Z - R relationship. The relative improvements range from 7
to 34 % depending on model and performance measure. The measures are even
better than the Z - R relationship retrospectively optimized for this very location. |
|
|
|
|
|
|