1. Christoph Schraff Deutscher Wetterdienst, Offenbach, Germany
KENDA operationalisation & status
Christoph Schraff Deutscher Wetterdienst, Offenbach, Germany
Contributions / input by:
Hendrik Reich, Andreas Rhodin, Roland Potthast, Klaus Stephan, Ulrich Blahak,
Michael Bender, Elisabeth Bauernschubert, Axel Hutt, … (DWD)
Daniel Leuenberger, Alexander Haefele (MeteoSwiss); Sylvain Robert (ETH)
Chiara Marsigli, Virginia Poli, Tiziana Paccagnella, Thomas Gastaldo (ARPA-SIM)
Lucio Torrisi, Francesca Marcucci, Valerio Cardinali (COMET)
Mikhail Tsyrulnikov, Dmitri Gayfulin (HMC)

2. KENDA operationalisation & status KENDA-O overview
PP KENDA-O : Km-Scale Ensemble-Based Data Assimilation
for the use of High-Resolution Observations
(Sept – Aug. 2020)
Task 1: further development of LETKF scheme
(work towards operationalization: additive covariance inflation)
limiter to soil moisture perturbations
AMDAR bias correction, RTPS, AMDAR RH
Task 2: extended use of observations
Mode-S aircraft (afternoon or tomorrow)
GPS slant total delay (afternoon)
Task 3: lower boundary: soil moisture analysis using satellite soil moisture data
Task 4: adaptation to ICON-LAM, hybrid methods / particle filters

3. operational use of KENDA
MeteoSwiss: KENDA operational for EPS
since 19 May (x = 2.2 km)
KENDA for det. COSMO-1: 2018
(T,qv in PBL still worse, some problems
 slides)
DWD: KENDA operational for det. + EPS
since 21 March (x = 2.8 km)
ARPAE-SIMC: KENDA operational for determ. forecast
(Italy) since May 2017, for EPS soon (2.2 km)
COMET: KENDA code adapted to include
(Italy) required capabilities of COMET system
and run in a parallel suite (x = 10 km)
(still slightly worse than COMET code)

4. KENDA-LETKF pre-operational setup
KENDA at DWD:
KENDA-LETKF pre-operational setup
 KENDA: 4D-LETKF + LHN (latent heat nudging for assimilation of radar precip)
benchmark:
operational Nudging + LHN
operational EPS
(40 mem.)
 K: Kalman Gain
for ensemble mean
(unperturbed)
1 hour
1 hour
(pre-) operational settings ( Schraff et al. 2016, QJRMS) :
adaptive horizontal localisation (keep # obs constant, 50 km ≤ s ≈ std dev ≤ 100 km)
adaptive mutliplicative covariance inflation (obs-f.g. statistics) RTPP (p = 0.75)
explicit soil moisture perturbations
conventional obs types only (radiosonde, aircraft, wind profiler, synop)

5. pre-operational KENDA suite at DWD in summer 2016:
effect of soil moisture perturbations (SMP)
24 Aug. 2016
soil moisture layer 5,
diff. betw. 2 mem. (1,8)
T2m at noon,
diff. betw. 2 mem. (1,8)
200
100
-100 10 5 -5
-10
standard soil moisture perturb.:
 T2m deviations of individual
ensemble members
unrealistically large
in some situations
100 50
-50 10 5 -5
-10
test: soil moisture perturbations
reduced by 50 % :
 T2m deviations realistic
 implement limiter to spread
of soil moisture index (SMI)
and assess impact on LETKF
(spread)

6. description of new setup with limiter for soil moisture perturbations
current setup: soil moisture perturbations:
 = SMI / day
modifications for parallel suite:
no explicit SM perturbations added
where spread(SMI) > 0.15
 “soft limiter” for SMP
depth weight of perturbations
explicit soil temperature (T_SO) perturbations:  = 0.48 K / day

7. soft limiter of soil moisture perturbations:
impact on a sunny day (23 June 2016)
spread of
soil moisture index
(14 UTC)
soil level 2 : 1 – 3 cm
soil level 5 : 27 – 81 cm
original
(Exp 10396)
with SMP limiter (Exp 10416)
soft limiter: SMI spread limited to ~ 0.15 SMI

8. soft limiter of soil moisture perturbations:
impact on a sunny day (23 June 2016)
spread of
soil temperature
soil layer 1: 0 – 1 cm
afternoon (14 UTC)
night (0 UTC)
original
(Exp 10396)
with SMP limiter (Exp 10416)
T_SO spread decreased mostly at daytime

9. soft limiter of soil moisture perturbations:
impact on a sunny day (23 June 2016)
spread of
2-m temperature
afternoon (14 UTC)
mid-level cloud (12 UTC)
Meteosat (12 UTC)
original
(Exp 10396)
with SMP limiter (Exp 10416)
T_2M spread decreased mainly in sunny areas

10. soft limiter of soil moisture perturbations:
impact on spread of T2M, day by day
spread of 2-m temperature, time series of 0-UTC runs
26 May – 27 June 2016
afternoon (14 UTC)
sunny
day
– 25 %
reduction 5 – 25 %
1 – 25 June 2017
many sunny days
reduction 10 – 35 %
T_2M spread reduction larger, due to sunny days, drier soil

11. soft limiter of soil moisture perturbations:
impact on spread of T2M, day by day
spread of 2-m temperature, time series of 0-UTC runs
26 May – 27 June 2016
night (2 UTC)
reduction ≤ 10 %
1 – 25 June 2017
reduction 5 – 18 %
T_2M spread reduction larger (as for soil temperature, effect from daytime?)

12. soft limiter of soil moisture perturbations:
impact on spread, surface verification
spread
0-UTC runs
26 May – 27 June 2016
original
SMP limiter
1 – 25 June 2017
parallel suite: spread more strongly reduced, but still larger than in old period

13. why is the reduction of spread of T2M larger than in previous test?
relative difference [%] of domain-averaged soil moisture
in deterministic run (on 22 June 2017/16)
(SM2017 – SM2016)/SM2016
 soil is drier
in 2017 period

14. /(SMS_new_2017 + SMS_new_2016)*2
why is the reduction of spread of T2M larger than in previous test?
scaled differences [%] of domain-averaged spread of soil moisture (SMS)
on 22 June (“2017” test)
resp. 22 June (“2016” test)
(SMS_2017 – SMS_2016)
/(SMS_new_ SMS_new_2016)*2
2017
2016
(SMS_old – SMS_new)
/(SMS_new_ SMS_new_2016)*2
SMS limiter results in larger
spread reduction in 2017
“old” setting: operational
“new” setting: with soft limiter
 spread larger in 2017,
particularly in old setting 14

15. reduction of [%] of rmse
impact of SMP limiter on scores, deterministic:
surface verification
SMP limiter vs. original
T 2M
RH2M ps
26 May – 27 June 2016
reduction of [%] of rmse
T 2M
RH2M ps
1 – 25 June 2017
forecast lead time [h]
parallel suite: still neutral (slightly better for RH2M in first 4 hours)

16. (averaged over lead times & initial times)
impact of SMP limiter on scores, deterministic:
radiosonde verification
26 May – 27 June 2016 T RH
wind
speed
wind direct.
1 – 25 June 2017
pressure [hPa] T RH
wind
speed
wind direct.
reduction [%] of rmse
(averaged over lead times & initial times)
parallel suite: neutral

17. impact of SMP limiter on scores, deterministic:
precipitation against radar
0-UTC runs
12-UTC runs
original
SMP limiter
26 May – 2 July
2016
1-hrly precip
FSS
( 30 km )
1 mm/h
1 – 12 June 2017
parallel suite: no improvement , neutral

18. impact of SMP limiter on scores, EPS:
surface verification (1 – 25 June 2017)
spread
T 2M
RH2M ps
original
SMP limiter
RMSE
CRPS
reduction [%] of CRPS

19. impact of SMP limiter on scores, EPS:
radiosonde verification (1 – 25 June 2017)
reduction [%] of CRPS
orig
with SM
limiter

20. impact of SMP limiter on scores, EPS:
radiosonde verification (1 – 25 June 2017)
reduction [%] of CRPS
orig
with SM
limiter

21. impact of SMP limiter on scores, EPS:
radiosonde verification (1 – 25 June 2017)
reduction [%] of CRPS
orig
with SM
limiter

22. Brier Score decomposition:
impact of SMP limiter on scores, EPS:
precip / radar verif. (26 May – 12 July 2016)
0-UTC runs
slightly better reliability (except high threshold) better resolution (not susceptible to calibration!)
lead time [h]
1 mm/h
0.1 mm/h
5 mm/h
the lower the better
the higher the better
reliability:
average agreement betw. fcst & obs val.; related to (cond.) bias
original
SMP limiter
Brier Score decomposition:
 resolution
 reliability
threshold [mm/h]
resolution:
ability of forecast to separate
one type of outcome from another

23. impact of soft SMP limiter: summary
impact of reduced soil moisture perturbations on scores:
T_2M spread strongly reduced in sunny days as required
T_2M + RH_2M spread moderately reduced in general
in June 2017 more than in June 2016, because in 2017:
more sunny days
original soil moisture perturbations larger
drier soil
deterministic scores: neutral
EPS: slightly (more) reduced (bias) errors (upper-air T, surface pressure), slightly improved resolution for precip
hypothesis: due to introduction of additive covariance inflation (in February), reduced soil moisture perturbations can be afforded or are even beneficial
 soft limiter to soil moisture perturbations operational since early July

24. Task 1, tests for refinements: bias correction for AMDAR
global bias correction (BC) for AMDAR depends on:
individual aircraft
flight phase
time: online bias correction (time scale ~ 30 days (?))
tests (2 weeks):
use BC values from global system
online BC in KENDA, starting from global values
 impact similar to 1), due to large time scale (which should be even larger than for global system, due to less data)
online BC in KENDA, starting from zero  little impact, due to large time scale
online BC in KENDA would require very large time scale or could become statistically ‘unstable’, would require very long experiments
caveat: AMDAR obs biased towards global ICON system

25. bias correction for AMDAR
Task 1, tests:
bias correction for AMDAR
forecast verification (26 May – 10 June 2016)
wind dir. RH
wind
dir.
wind speed ps T
wind
speed
RH2M
T2M
TD2M
change in RMSE [%]
positive impact from bias-correction

26. bias correction for AMDAR
Task 1, tests:
bias correction for AMDAR
26 May – 10 June 2016
BC-AMDAR vs. ref
0-UTC runs
6-UTC runs
1-hrly precip
FSS (30 km)
improvement
1 mm/h
12-UTC runs
18-UTC runs
mean / 5% / 95%
confidence interval
(bootstrapping)
AMDAR bias correction: mixed impact

27. Task 1, tests for refinements:
AMDAR RH data :
very small, neutral impact
Relaxation To Prior Spread (RTPS):
replacing:
adaptive multiplicative covariance inflation (based on obs – fg statistics)
Relaxation To Prior Perturbations (RTPP)

28. Task 1, tests: RTPS RH T forecast verification (26 May – 10 June 2016)
wind dir. RH
wind
dir.
wind speed ps T
wind
speed
RH2M
T2M
TD2M
change in RMSE [%]
slightly negative / neutral impact

29. Task 1, tests: RTPS 0-UTC runs 12-UTC runs 0.1 mm/h 1 mm/h
26 May – 10 June 2016
0-UTC runs
12-UTC runs
Ref
RTPS
0.1 mm/h
1-hrly precip
FSS
( 30 km )
1 mm/h
RTPS: clear long-lasting positive impact on 0- / 12-UTC runs

30. Task 1, tests: RTPS 6-UTC runs 18-UTC runs 0.1 mm/h 1 mm/h
26 May – 10 June 2016
6-UTC runs
18-UTC runs
Ref
RTPS
0.1 mm/h
1-hrly precip
FSS
( 30 km )
1 mm/h
RTPS: negative impact on 6- / 18-UTC runs for 1 mm/h

31. Task 2: Extended use of observations: ongoing
Progress in KENDA-O:
Task 2: Extended use of observations: ongoing
radar radial winds: DA exp., promising results
radar reflectivity: DA sensitivity tests, results mostly preliminary:
 ARPAE-SIMC (≤ 4 radars): 1h / 30’/ 15’ cycling, temporal thinning, RTPS, addit. infl.
 DWD: warm bubbles, remapping of Z for Gaussianity, 1.1km, 2-moment microphys.
GPS slant total delay: technical work incl. monitoring, in V5.04d, new DA exp. running
(old exp: benefit on precip, elsewhere mixed (negative at low levels))
SEVIRI WV: many (mostly clear-sky) sensitivity DA tests, inconclusive, no benefit yet
SEVIRI cloud top height: no resources
T2M, RH2M: delayed (resources at MCH in 2018)
Mode-S aircraft : DA exp., promising results
Raman lidar (T-, q- profiles): quality of Payerne obs improved, DA being set up
AMSU, ATMS, IASI clear-sky radiances: no resources
3 new projects at DWD ( WG1): SEVIRI VIS, lightning, (nowcast) objects

32. Task 2, extended use of observations: Mode-S aircraft
derived from radar data from air- traffic control,
processed + provided by KNMI
wind vector + temperature,
T derived from Mach number
every 4 sec
compared to AMDAR:
>10 x more data
no humidity, larger T error
from: Lange and Janjic, MWR 2016

33. Task 2, extended use of observations: Mode-S aircraft
bug fixes in COSMO V5.04d
experiment 26 May – 10 June (‘E19’)
best results with thinning (40 % active)
Mode-S dominate above 800 hPa
number of active obs
a-posteriori Desroziers statistics from DA experiment for estimation of obs errors:
unexpected large Mode-S wind errors
data processing corrected by KNMI
since 15 May 2017,
re-processed historical data on request
new exp. ‘E22’ with re-processed data:
Mode-S wind obs errors similar to AMDAR
(but specified obs error variance still large)

34. Task 2, extended use of observations: Mode-S aircraft
radiosonde verif. for DA cycle
(26 May – 10 June 2016)
positive impact from Mode-S, enhanced by corrected data

35. Task 2, extended use of observations: Mode-S aircraft
Mode-S aircraft : forecast verification
(26 May – 10 June 2016)
wind dir. RH
wind
dir.
wind speed ps T
wind
speed
RH2M
T2M
TD2M
change in RMSE [%]
positive impact from Mode-S throughout

36. Task 2, extended use of observations: Mode-S aircraft
26 May – 10 June 2016
0-UTC runs
12-UTC runs
Ref
Mode-S (E19)
Mode-S (E22)
0.1 mm/h
1-hrly precip
FSS
( 30 km )
1 mm/h
Mode-S: clear long-lasting positive impact

37. trial in parallel suite
Task 2, extended use of observations:
Mode-S aircraft
1 – 28 Aug. 2017
trial in parallel suite
1 – 20 Aug , 0-UTC runs RH ps
0.1 mm/h T
RH2M
1 mm/h
wind
speed
T2M
change in RMSE [%]
Aug. 2017: much smaller positive impact than in 2016 convective period ! Why?
weather? data quality?
data cut-off time? (the last 15 min of Mode-S obs not available in time)

38. Task 2, extended use of observations: Mode-S aircraft
trial in parallel suite
data availability at cut-off time = 15 min !
AMDAR obs valid -29 to -15 min
AMDAR obs valid -14 to -0 min
70 % available
Mode-S obs valid -29 to -15 min
97 % available
95 % available
Mode-S obs valid
-14 to -0 min
0 % available

39. Task 4.1: KENDA for ICON-LAM (incl. EnVar)
MEC-based LETKF for ICON-LAM: ready to be tested (e.g. grid pt. assignment)
(with COSMO obs operators) except for (hydrostatic) balancing step(s)
& exclusion of obs near lateral BC
full 4-D LETKF for ICON-LAM:
implementing part of DACE code, incl. COSMO + global obs operators in ICON,
 first working version Nov. 2017
thereafter: complement global obs operators by functionality of COSMO operators
 needed for EnVar (TL + adjoint of obs operators, incl. radar etc.)
first tests LETKF with ICON-LAM planned for spring (Hollborn)

40. Task 4.2: Particle Filter (PF) methods
(to account for non-Gaussianity)
hybrid ETKPF (Sylvain Robert (ETH) et al., study finished):
1-week DA with PF for conv. obs in COSMO: f.g. (+1h) slightly improved over LETKF
studying several algorithms for adaptive choice for weight of EnKF and PF in analysis
(idea: use PF where useful, fall back on EnKF where Gaussian or PF does not work well)
QJRMS: “A local ensemble transform Kalman particle filter for convective scale DA”
EnVar
PF-Var
global TEMP verif, 2 – 24 May.2016
hybrid PF-Var: using PF ensemble in EnVar
for deterministic analysis:
 forecasts only slightly worse than
operational EnVar with ensemble B from LETKF
(promising!)
Local adaptive PF (Roland Potthast et al., ongoing):
work with global ICON
(improved resampling to create new members)
PF runs stably in 1-month tests
scores still about 20% worse than LETKF

41. Task 2: Extended use of observations
Progress in KENDA-O:
Task 2: Extended use of observations
GNSS slant total delay (STD, Michael Bender)
experiments: positive impact on precip so far,
mixed otherwise (negative at low levels)
STD observation operator in COSMO V5.04d
AutoAlert implemented (for automatic QC of each station & product, bías correction
files, monitoring of station coord., visualisation, … )
plans:
: refinement of DA, impact experiments,
make STD technically ready for operational use, passive monitoring
(converter ASCII (NetCDF) to BUFR  NetCDF : reader in COSMO)
2018: operational introduction of STD in KENDA

42. Task 2, extended use of observations: GNSS Slant Total Delay (STD)
Monitoring of STD obs from GFZ with new ‘EPOS-8’ processing
bias: (rather) small
relative std. dev. (std.dev./mean-abs-value):
does not increase with decreasing elevation
STD data quality does not decrease
with decreasing elevation (  7°)

43. Task 2, extended use of observations: GNSS Slant Total Delay (STD)
Local Ensemble Transform KF  obs location & localisation of STD’s / ZTD’s

44. Task 2, extended use of observations: GNSS Slant Total Delay (STD)
for specific humidity
analysis increments:
analysis increment
for a single (STD / ZTD) obs  a number : c

45. Task 2, extended use of observations: GNSS Slant Total Delay (STD)
analysis increment
for temperature
analysis increments:
for a single (STD / ZTD) obs  a number : c

46. Task 2, extended use of observations: GNSS Slant Total Delay (STD)
analysis increment
for pressure
analysis increments:
for a single (STD / ZTD) obs  a number : c

47. Task 2, extended use of observations: GNSS Slant Total Delay (STD)
analysis increments
pressure temperature specific humidity
approx. optimal localisation scale ?
optimal to treat
STD/ZTD obs
as if located
approx. here (?)
possible implementation of individual obs re-location:
HT available in COSMO: approximate roughly &
store in ‘fof’ feedback file
(‘plevel’, ‘plev_width’)
Pf implicitly available in LETKF:
use model columns above station from ensemble to compute variance
(neglecting horizontal variability of variances)

48. Task 2: Extended use of observations

49. Task 2: Extended use of observations
Screen-level obs
RH2M (+ T2M) : tests at MeteoSwiss in 2018
10-m wind : new criteria (dep on. roughness length + d2zoro/(dx2dy2))
for station selection of 10-m wind in COSMO V5.04d
Ground-based remote sensing (T + qv prof.: Raman lidar, MW radiometer)
MeteoSwiss: improved quality of temperature obs from Payerne Raman lidar
setting up DA of 0.5-hourly lidar qv + T as TEMP bulletin

50. Task 2: Extended use of observations
SEVIRI radiances
direct use of all-sky (cloudy) SEVIRI WV + IR window radiances (Axel Hutt)
reproduce method of Harnisch et al. 2016, QJ, for cloud-dep. obs errors + bias correction
technical work (upgrade to V5.04d, RTTOV-12, BACY, …)
sensitivity experiments with use clear-sky WV channels:
data selection (clear-sky; high orography),
obs operator (effective radius related to qc, qi), obs error, bias correction, thinning, …
 results mostly a bit inconclusive, bias correction is needed,
overall impact of clear-sky obs neutral or negative
LMU: new PostDoc Josef Schröttle
use of NWC-SAF cloud top height (CTH) product: currently no resources
(outside KENDA-O:) direct use of all-sky (cloudy) SEVIRI VIS/NIR radiances
Lilo Bach at DWD since ~ July (work together with Leonard Scheck, LMU)