[Mitgcm-support] c20000630 vs kf023b3, kf024f, kf024i, kf025e, kf026a, and kf029

Wed Jul 9 15:56:17 EDT 2003

HyperNews notification of new message.  See: 
  http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/assim/106/2/3.html

Comparison of Kalman filters kf023b3, kf024f, kf024i, kf025e, kf026a,
and kf029 with control integration (c20000630)

A full description of Kalman filter estimates is here[1] . Brief summary
of selected estimates follows:

  kf022e  : BT filter, calibrated Q
  kf023b3 : BC filter, NCEP Q
  kf024f  : BC filter, calibrated Q.
  kf024i  : BC filter, Same as kf024f but with tapered P
            towards NS open boundary.
  kf025e  : BC-BT combined filter, calibrated Q.
  kf026a  : BC-BT combined filter, NCEP Q.
  kf029   : BC-BT filter combining "P"s of kf024i and kf022e. 

Profiles of mean temperature, standard deviation temperature, standard
deviation difference temperature from TAO data, and correlation
coefficient with TAO data are here[2] . Overall, all Kalman filter
estimates have more or less the same skill in the top 200 m of the water
column and their varibility is significantly closer to that of TAO data
than the control integration c20000630. Below 200 m, the estimates with
NCEP Q (kf023b3 and kf026a) appear to do a better job than the estimates
with calibrated Q (kf024f, kf024i, kf025e, and kf029).

Maps of simulation/assimilation skill and correlation versus TAO mooring
temperature follow:

 kf023b3[3]
 kf024f[4]
 kf024i[5]
 kf025e[6]
 kf026a[7]
 kf029[8]

Results are shown at 3 depths for each Kalman filter estimate, 75, 150,
and 300 m. The top three panels represent skill defined as
"(var(simulation-TAO)-var(assimilation-TAO))/var(TAO)". The bottom three
panels represent improvement in correlation,
"cor(assimilation,TAO)-cor(simulation,TAO)". In each case reds and
yellows indicate regions where the assimilation is closer to TAO data
than the simulation, while blues indicate regions where assimilation is
further away from TAO than simulation. Locations of TAO moorings are
indicated by dots. Overall, the figures show that all Kalman filter
estimates are qualitatively the same, but that the estimates with NCEP Q
(kf023b3 and kf026a) are closer to TAO data at 300 m.

A final map[9] compares kf023b3 (Kalman filter with NCEP Q with the most
skill relative to TAO data at 300 m depth) and kf029 (Kalman filter with
calibrated Q with the least skill relative to TAO data at 300 m depth).
Again top three panels represent skill of each filter estimate relative
to TAO data and bottom three panels represent correlation. Reds and
yellows are areas where kf023b3 is closer to TAO data and blues indicate
regions where kf029 is closer to data. Degradation of kf029 relative to
data, as compared to kf023b3 is most pronounced at 300-m depth in a
broad region centered about the equator and the date line.

The story from velocity comparisons is less clearcut. Follow summary
profiles for zonal and meridional velocity at each of three TAO current
meter locations, 165E, 140W, and 110W, and their mean.

 165E zonal[10]
 165E meridional[11]
 140W zonal[12]
 140W meridional[13]
 110W zonal[14]
 110W meridional[15]
 mean zonal[16]
 mean meridional[17]

Overall, Kalman filter zonal velocity estimates have about 40 percent
more energy than the simulation, but still much less than TAO current
meter moorings. They are slightly more correlated with TAO data than the
simulation above 200 m. Meridionally, there is little change on average.

At 165E, Kalman filter zonal velocity estimates fare better than control
integration above 50 m depth, but worse below 50 m. One surprising
conclusion is that the model is highly correlated (.8) with TAO data
below 200 m depth, but that the Kalman filter estimates much degrade
that correlation (0-0.2). Kalman filter meridional velocity fares
slightly better with overall improvement relative to simulation down to
150 m depth. Below that depth, estimates with NCEP Q (kf023b3 and
kf026a) continue to be closer to TAO current meter data, but the
remaining estimates are slightly degraded.

Estimates with NCEP Q (kf023b3 and kf026a) can also be singled out at
the other locations. On average they fare slightly worse than the other
estimates below 100 m depth at 140W and 110W, but slightly better than
the other estimates above 50 m depth at 110W.
-----------------------------------------------------------------------
[1] http://escher:2000/hosts/escher/escher6/medea/btang/kf_bc_runs/kf_run.txt
[2] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/global_profiles.ps
[3] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf023b3_map.ps
[4] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf024f_map.ps
[5] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf024i_map.ps
[6] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf025e_map.ps
[7] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf026a_map.ps
[8] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf029_map.ps
[9] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/TempStats/kf023b3_kf029.ps
[10] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/165e_zonal_prof.ps
[11] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/165e_merid_prof.ps
[12] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/140w_zonal_prof.ps
[13] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/140w_merid_prof.ps
[14] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/110w_zonal_prof.ps
[15] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/110w_merid_prof.ps
[16] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/mean_zonal_prof.ps
[17] http://escher:2000/hosts/triton/dm1/dimitri/data/tao/matlab/FIGS/Currents/mean_merid_prof.ps