1. PP INSPECT is finished, but the work is going on
Anastasia Bundel
RHM
COSMO GM in Saint Petersburg, 04 September 2018

2. INSPECT final report
The report is prepared and sent to the SMC members for eventual comments, critics, or suggestions. No comments so far.
When it’s approved, it will be published on the COSMO web-site.

3. Feedback from MesoVICT
BAMS paper: Dorninger, M., E. Gilleland, B. Casati, M. Mittermaier, E. Ebert, B. Brown, and L. Wilson, 2018: The set-up of the Mesoscale Verification Inter-Comparison over Complex Terrain (MesoVICT) Project. Bull. Amer. Meteor. Soc. doi: /BAMS-D , in press.

4. Feedback from MesoVICT, cont.
Regular participation of COSMO people in MesoVICT webinars once a month
MesoVICT final meeting 2019, variants:
- ICAM International Conference on Alpine Meteorology
- COSMO GM in Rome: “since the COSMO community has been so active in the project, having it joint with the COSMO community would be a fitting acknowledgement of COSMO contribution” (from organizers’ letter)
- Special meeting

5. One of the INSPECT outcomes:
Several ways of compact visualization of neighborhood, CRA, and SAL methods (DWD, MCH, RHM, HNMS).
Especially for neighborhood scores, such a cumulative framework can be implemented as part of COSMO Verification activity (possibly using VAST)

6. Processing large volumes of data: CRA
for STEPS nowcasting at Roshydromet (Anatoly Muraviev)
The core of the system is the statistical STEPS scheme (Short Term Ensemble Prediction System) (Bowler N. et al., 2006)
Verification period: May-September 2017
9 radars in Central Russia
Forecasts for intense precipitation areas only were analyzed (169 situations)
10 min time step until 3 h
Grid size of about 2 km
SpatialVx was used to indentify objects and to calculate CRA scores.
The objects with areas less than 35*35 grid points (about 70х70 km) and lager than 128х128 grid points (about 250х250 km) were excluded from analysis using min.size and max.size option in FeatureFinder function. The radius of averaging for convolution smoothing was chosen empirically as 9 grid points (18 km)
Precipitation threshold for object recognition showed here is 1 mm/h

7. SpatialVx: first и last objects, 17 May 2017
RADAR
STEPS
RADAR
STEPS
The radar object splitted in two
by the end of the forecast period 7

8. R-SpatialVx, CRA STEPS forecast: Radar RAKU 20170517_1120
Such CRA tables are aggregated over the whole period
May-September 2017 8

9. Statistics of CRA object centroid longitude shift: mean, median, quartiles, max, and min.
Critical shift value empirically defined as 35 grid points (radius of a smallest round object)
Red : shifts for all objects don’t exceed the critical value
Green : shifts for not less than 50% of objects don’t exceed the critical value (until 90 min lead time)
Systematic shifts and other CRA error components can be determined in such a way.

10. Analysis of object sizes
Analysis of object sizes. Extreme value analysis: Introductory study Motivation: to understand what are our objects (features/contiguous areas/entities) statistically

11. Distribution histograms of object sizes
One radar (RAKU) data
Number of objects (>=1 mm) in winter is twice less than in summer, but the relative frequencies of sizes are comparable
Upper row for warm period,
Lower row, cold period
Red : number of objects,
Blue: relative frequencies
OBJECT SIZE in GRID POINTS*1000

12. Estimating the distribution of extreme values :
R fevd module
R extRemes package( E.Gilleland)
Module fevd – fitting extreme value distributions to data, plotting histograms, parameter estimation, distribution densities, fitting of tails, quantile-quantile (QQ) plots, statistical significance, …
Distributions and parameters:
GEV – Generalized Extreme Value distribution is a family of continuous probability distributions developed within extreme value theory to combine the Gumbel, Fréchet and Weibull families
GP – Generalized Pareto
PP – Poisson Process
Gumbel: a type of GEV
Exponential
Methods for estimating parameters:
MLE (Maximum Likelihood Estimation),
GMLE (Generalized MLE)
Bayesian
Lmoments
Some verification specifics

13. Gumbel distribution doesn’t fit well, as well as Poisson (not shown)
Fitting object size distribution using fevd : QQ-plots and probability density functions (pdf) for warm period (RAKU radar) for a threshold of 1600 grid points
Heavy tail
Heavy tail
Light tail
Light tail
Some verification specifics
Gumbel distribution doesn’t fit well, as well as Poisson (not shown)
Q-Q plots: GEV and GР-Pareto, acceptable up to 13000 grid points area
Here, the serial correlation (development of the same object in time) was not taken into account. Now, the study is ongoing to decimate the series of forecasts based on mesoscale time-space structure

14. What was not completely fulfilled within INSPECT and needs development
Explicitly introducing orography factor
Applications to wind and other variables besides precipitation will be continued
Focus on HIW and more user-specific variables, exploration of non-standard observations – link to WG4 and WG7!
Further applications to ensembles and introducing observation uncertainty
Processing large volumes of rapidly updating data from NWP and nowcasting systems (“terabytes of data”), calculation efficiency will be of utmost importance!

15. Plans
As this is a cross-cutting approach among different working groups, maybe, it is not necessary to initiate a special PP about the spatial methods, but better to have dedicated tasks in new joint projects, in the first place, the project focusing in HIW.
INTERP and INSPECT knowhow is used in COSMO
Research on spatial methods is going on
Conclusion

16. THANK YOU FOR YOUR ATTENTION!

17. Distribution parameters of object sizes (1 mm/h) identified by FeatureFinder (size is the number of contiguous grid points)
R-summary:
min, q25, med, mean,q75,max,IQR,std
1600 grid points is higher than q75 for all radars ->
Threshold of 1600 grid points was chosen for extreme value analysis
Some verification specifics
Different radars 17