Verification in Complex Terrain with Ensemble-Analysis including observation uncertainty
Simon Kloiber, University of Vienna
Traditionally, verification means to verify a forecast (ensemble) with the truth represented by observations. The observation errors are quite often neglected arguing that they are small when compared to the forecast error.
In this study as part of the MesoVICT (Mesoscale Verification Intercomparison over Complex Terrain) project it will be shown, that observation errors have to be taken into account for verification purposes. The Observation uncertainty is estimated from the VERA (Vienna Enhanced Resolution Analysis) and represented via two analysis ensembles which are compared to the forecast ensemble. For the whole study the COSMO-LEPS provided by Arpae-SIMC Emilia-Romagna are used as forecast ensemble. The time period covers the MesoVICT core case from 20-22 June 2007.
For the first part, developing verification methods to compare two ensembles, knowledge about their distribution is necessary. To answer that question, tests have been executed (Kolmogorov-Smirnov-Test, Finkelstein-Schafer Test, Chi-Square Test etc.) and they haven’t shown an exact mathematical distribution. So the main focus is on non-parametric statistics (e.g. Kernel density estimation, Boxplots etc.) and also the deviation between “forced” normal distributed data and the kernel density estimations.
In the second part, the observational deviations due to the analysis ensembles are represented. A boxplot shows the range of a score, which is calculated multiple times with every ensemble member as the “true” observation. Additionally, the bootstrapping method may also provide a boxplot, if two ensembles getting compared. These possible approaches to incorporating observational uncertainty into the computation of statistics will be discussed in the talk.
This seminar will be webcasted at: http://www.fin.ucar.edu/it/mms/fl2-live.htm
Monday, December 12, 2016 | 2-3 pm | FL2 - Room 1001
National Center for ATmospheric Research, Boulder