1. Assimilation of weather radar observations at the UK Met Office
David Simonin
6th WMO workshop on the Impact of Various Observing Systems on NWP
10-13 May, 2016

2. Acknowledgements Met-Office radar team
Lee Hawkness-Smith (Reflectivity)
Bruce Macpherson and Gareth Dow (UK4)
Jo Waller (Correlated observation error)
John Nicol (Refractivity)

3. Content: Motivation UK radar network and observations
Impact on forecasts
Doppler wind
Doppler wind + LHN of rain rate
Direct assimilation of reflectivity
How to increase positive impact QC
Observation operator
Observation Error
Conclusion

4. Motivation
The development of operational convective-scale NWP (e.g. NWP-based nowcasting system) models that represent convection explicitly require the development of high resolution data assimilation systems to provide initial conditions at the appropriate scale.
Radar data are the “perfect” candidate for assimilation:
High resolution
High repetition time
Large area coverage

5. Radar network and observations

6. UK Radar Network 18 radars currently in the network PPI Scan
Introduction
18 radars currently in the network
PPI Scan
5 elevations between 0˚ and 9˚
1˚ Ray, 75/300/600 m Gate
Volume scan available every 5 minutes
Format: ODIM HDF5 (OPERA)
One weather radar ~10 million obs. per hour reflectivity, refractivity and Doppler wind

7. Two modes of acquisition
UK Radar Network
Two modes of acquisition
Long Pulse acquisition
600m / 250km
Short Pulse acquisition
75m / 100km
Reflectivity
Doppler Wind
Phases
LHN of rain rate
Doppler wind
Operationally assimilated
Direct assimilation
Doppler wind with correlated obs error
Refractivity change
Active research

8. Impact on forecast

9. Impact on forecast Doppler Radial wind NDP domain (1.5km)
3DVAR, Hourly cycle, nested in UKV.
70 Levels
LBCs updated every 30mins; refreshed every 6 hours
Super-observations, 1 volumes scan (analysis time)
Example: T+2 rain rate

10. Impact on forecast Doppler Radial wind
Aggregated RMSE – Wind profiler (u-comp.)
Aggregated RMSE – Doppler wind

11. Aggregated FSS using hourly rain accumulation
Impact on forecast
Doppler Radial wind
Aggregated FSS using hourly rain accumulation

12. Doppler Radial wind + LHN of rain rate
Impact on forecast
Doppler Radial wind + LHN of rain rate
UK4 (4km) model
LBCs provided by NEA (global from 2012)
3DVAR, 3 hourly
Impact UK index (1 month) - Doppler wind and LHN rain rate
Weighted Basket of Indices / Combo of ETS & RMS scores
Vis
Precip
Cloud Cover
Cloud Base
Height
Temp
Wind
Overall
-0.006
+0.237
+0.022
+0.018
+0.048
+0.091
+1.64%
Courtesy of Bruce Macpherson and Gareth Dow

13. Reflectivity Observations
Impact on forecast
Reflectivity Observations
Direct assimilation of reflectivity observation
4Dvar hourly cycle / UKV (No outer-loop)
3 volumes scans (0, 15, 30 minutes)
Both dry and wet observations are used.
Radar
control
Experiment
In the case of spurious or excessive precipitation, the reflectivity assimilation scheme is beneficial (but the effect is small).
Overall, there is a strong dry bias.
Assimilation dominated by dry observations.
Reduce weight
More thinning
Scheme affected by PF model limitations.
introducing outer-loop
The key issue to be addressed before reflectivity assimilation can be considered a candidate to replace latent heat nudging is the dry bias of the assimilation scheme. The dry bias may simply be due to the fact that the vast majority of reflectivity observations are dry, and thus the overall weight of the dry observations exceeds that of observations of precipitation. Thus Var may find a better fit to the reflectivity observations by reducing precipitation everywhere than by fitting the relatively small fraction of observations of precipitation.
Generating cloud where none previously exists is not possible in a linear model such as the PF model. The use of multiple outerloops in Var would help reduce this problem, as the full non-linear forecast model is able to generate cloud
Ideas to investigate for addressing this issue include:
thinning the dry observations more than rainy observations,
assigning a larger observation error to dry observations than rainy observations,
using an asymmetrical Huber norm, where a larger weight is allowed for positive innovations than for negative innovations
Courtesy of Lee Hawkness-Smith

14. How to increase positive impact

15. Impact from Quality control
Example: Doppler radial wind gross error
Radar failure
Sea clutter
Unfolding
As John show in the opening presentation of this workshop, Quality Control for satellite level 1/2 observation is key to the impact. The same is try for Radar data. Here is an example of problem that could occur when acquiring Doppler radial wind.

16. Impact from Quality control
Improvement in acquisition
Improvement in acquisition (dual polarisation) enables a better classification.
Old QC
New QC
Sea clutter black listed
Courtesy of Nawal Husnoo

17. Impact from Quality control
Retrieving Changes in Refractivity
Refractivity change retrieval depends on the assumption that a change in the phase results only from a change in the refractive index.
But it can result also from the motion of the target.
 We need good quality clutter
[A]
[B]
[C]
[D]
Refractivity change derived using a Quality index 
The quality index is based on the average coherence of the target and is used as a weight in the retrieval.
Refractivity change derived without the Quality index 

18. Impact from Quality control
Using additional parameters
Make use of extra parameters such as the spectral width.
Low spectral width:
Smooth flow
High spectral width:
Turbulent flow
Noise contamination
Doppler radial wind
Doppler Spectral width
Improve the Quality Control
Refine the observation error
Assign weight to the super-observation
etc...

19. Impact from the Observation Operator
The observation operator provides the link between the model variables and the observations.
It can be multi-variant and represent transformation(s) to go from model space into observation space.
Will influence:
O-B and O-A (fit to obs.)
Representativeness error
In practice:
How complex should it be?
Beam propagation, Beam broadening
Beam attenuation (due to hydrometeors), Hydrometeor fall speed (for radial velocities)
Should we use a retrieval?
Radar derived Refractivity’s changes retrieved from change in phase over dt and dx.
 The retrieval will introduce some correlated observation error
 Should we try to assimilate directly changes in phase? It is a lot harder!

20. Impact from the Observation Operator
Example: Doppler radial wind
Original Observation Operator
O-B RMSE over 1 month
Original H
New H
Introducing Beam Broadening
Reflectivity weighting (bright band)
Beam broadening to represent the cross section.
Improvements visible in the O-B
Reduced VAR’s performance:
Detrimental to the other wind observations.

21. Impact from the Observation Error
R = correlated error + uncorrelated error
Observation errors arise from four main sources:
Instrument error
Representativity error
Observation operator error
Observation pre-processing
Super-Observation:
It will reduce data volume and uncorrelated observation errors.
It will not remove the correlated part of the error.
Apply some thinning:
It could be used to reduce the effect of the correlated error.
Too much thinning could remove some convective-scale information.

22. Impact from the Observation Error
Example: Doppler radial wind
Methods to derive R:
Desroziers diagnostic (rely on O-B and O-A)  retrieve full R
Desroziers diagnostic:
σo2 is function of altitude. (Not gate size)
σo2 and correlations function of H and processing
Influence of H on R
Large reduction in length scale at far range (78km)
Slight reduction at near range (18km)
Operational H (triangles)
Improved H (diamonds).
Horizontal correlations larger than expected
Tests will start in the coming weeks.
Courtesy of Joanne Waller

23. Conclusion:
Convective-scale NWP has benefited from the assimilation of weather radar observations.
However, we still have a long journey ahead before we can make the most of it.

24. Conclusion: Convective scale NWP model are increasing in size:
 There is now a need for well design data exchange program.
Standardisation (meta data and additional variables)
Common file format
OPERA (a program from EUMETNET) is a good example.
Some important areas of research:
Direct reflectivity assimilation  replacement to LHN
Refractivity assimilation  will provide surface humidity information
Use of non-diagonal observation error matrix
Keep improving the observation quality

25. Thank you! Questions?