Introduction to data assimilation and least squares...
Transcript of Introduction to data assimilation and least squares...
![Page 1: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/1.jpg)
Introduction to data assimilationIntroduction to data assimilationand least squares methodsand least squares methods
Eugenia Kalnayand many friends
University of MarylandOctober 2008 (Part 2)
![Page 2: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/2.jpg)
Contents (1)Contents (1)• Forecasting the weather - we are really getting better!• Why: Better obs? Better models? Better dataassimilation?• Intro to data assim: a toy scalar example 1, wemeasure with two thermometers, and we want anaccurate temperature.• Another toy example 2, we measure radiance but wewant an accurate temperature: we will derive OI/KF,3D-Var, 4D-Var and EnKF for the toy model.
![Page 3: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/3.jpg)
Contents (2)Contents (2)• Review of toy example 1• Another toy example 2, we measure radiance but wewant an accurate temperature:• We derive Optimal Interpolation/Kalman Filter(sequential algorithms) for the toy model.• 3D-Var, 4D-Var (variational algorithms) and EnKF forthe toy model.• We compare the toy equations and the real equations• An example from JMA comparing 4D-Var and LETKF
![Page 4: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/4.jpg)
Typical 6-hour analysis cycle
Bayes interpretation: a forecast (the “prior”), is combined with thenew observations, to create the Analysis (IC) (the “posterior”)
![Page 5: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/5.jpg)
Intro. to data assim: toy exampleIntro. to data assim: toy example 1 summary1 summary
1
!a
2=1
!b
2+1
!o
2
If the statistics of the errors are exact, and if the coefficientsare optimal, then the "precision" of the analysis (defined asthe inverse of the variance) is the sum of the precisions ofthe measurements.
A forecast and an observation optimally combined (analysis):
Now we are going to see a second toy example of dataassimilation including remote sensing.
The importance of these toy examples is that the equationsare identical to those obtained with big models and many obs.
Ta= T
b+
!b
2
!b
2+!
o
2(T
o" T
b) with
![Page 6: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/6.jpg)
Intro. to Intro. to remote sensingremote sensing and data and dataassimilation: toy example 2assimilation: toy example 2
• Assume we have an object, a stone in space• We want to estimate its temperature T (oK) accurately but wemeasure the radiance y (W/m2) that it emits. We have an obs.model, e.g.:
y = h(T ) ! !T
4
![Page 7: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/7.jpg)
Intro. to remote sensing and dataIntro. to remote sensing and dataassimilation: toy example 2assimilation: toy example 2
• Assume we have an object, a stone in space• We want to estimate its temperature T (oK) accurately but wemeasure the radiance y (W/m2) that it emits. We have an obs.model, e.g.:
• We also have a forecast model for the temperatureT (t
i+1) = m T (ti)[ ];
e.g., T (ti+1) = T (t
i) + !t SW heating+LW cooling[ ]
y = h(T ) ! !T
4
![Page 8: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/8.jpg)
Intro. to remote sensing and dataIntro. to remote sensing and dataassimilation: toy example 2assimilation: toy example 2
• Assume we have an object, a stone in space• We want to estimate its temperature T (oK) accurately but wemeasure the radiance y (W/m2) that it emits. We have an obs.model, e.g.:
• We also have a forecast model for the temperature
• We will derive the data assim eqs (KF and Var) for this toysystem (easy to understand!)
T (ti+1) = m T (t
i)[ ];
e.g., T (ti+1) = T (t
i) + !t SW heating+LW cooling[ ]
y = h(T ) ! !T
4
![Page 9: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/9.jpg)
Intro. to remote sensing and dataIntro. to remote sensing and dataassimilation: toy example 2assimilation: toy example 2
• Assume we have an object, a stone in space• We want to estimate its temperature T (oK) accurately but wemeasure the radiance y (W/m2) that it emits. We have an obs.model, e.g.:
• We also have a forecast model for the temperature
• We will derive the data assim eqs (OI/KF and Var) for this toysystem (easy to understand!)• Will compare the toy and the real huge vector/matrixequations: they are exactly the same!
T (ti+1) = m T (t
i)[ ];
e.g., T (ti+1) = T (t
i) + !t SW heating+LW cooling[ ]
y = h(T ) ! !T
4
![Page 10: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/10.jpg)
Toy temperature data assimilation, measure radiance
We have a forecast Tb (prior) and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )
The new information (or innovation) is theobservational increment:
![Page 11: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/11.jpg)
Toy temperature data assimilation, measure radiance
We have a forecast Tb (prior) and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )
The new information (or innovation) is theobservational increment:
The final formula is the same as in example 1:
Ta = Tb + w(yo ! h(Tb ))
with the optimal weight w = !b
2H (!
o
2+ H!
b
2H )
"1
![Page 12: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/12.jpg)
Toy temperature data assimilation, measure radianceSummary for Optimal Interpolation/Kalman Filter (sequential):
Ta = Tb + w(yo ! h(Tb ))
with w = !b
2H (!
o
2+!
b
2H
2)"1
The analysis error is obtained from
!a
2= (1" wH )!
b
2
which can also be written as
1
!a
2=
1
!b
2+H
2
!o
2
"
#$%
&'
analysis
optimal weight
analysis precision=forecast precision + observation precision
![Page 13: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/13.jpg)
From a 3D-Var point of view,we want to find a Ta thatminimizes the cost function J:
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
![Page 14: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/14.jpg)
From a 3D-Var point of view,we want to find a Ta thatminimizes the cost function J:
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
This analysis temperature Ta is closest to both theforecast Tb and the observation yo and maximizes thelikelihood of Ta~Ttruth given the information we have.
It is easier to find the analysis increment Ta-Tb thatminimizes the cost function J
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
![Page 15: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/15.jpg)
From a 3D-Var point of view,we want to find a Ta thatminimizes the cost function J:
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
The cost function comes from a maximum likelihood analysis:
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
![Page 16: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/16.jpg)
From a 3D-Var point of view,we want to find a Ta thatminimizes the cost function J:
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
Likelihood of Ttruth given Tb: 1
2!"b
exp #(T
truth# T
b)2
2"b
2
$
%&
'
()
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
![Page 17: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/17.jpg)
From a 3D-Var point of view,we want to find a Ta thatminimizes the cost function J:
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
Likelihood of Ttruth given Tb: 1
2!"b
expTtruth
# Tb( )2
2"b
2
$
%&&
'
())
Likelihood of h(Ttruth) given yo: 1
2!" o
exp #h(Ttruth ) # yo( )
2
2" o
2
$
%&&
'
())
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
![Page 18: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/18.jpg)
From a 3D-Var point of view,we want to find a Ta thatminimizes the cost function J:
2Jmin
=(Ta ! Tb )
2
" b
2+(h(Ta ) ! yo )
2
" o
2
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
Likelihood of Ttruth given Tb: 1
2!"b
exp #Ttruth
# Tb( )2
2"b
2
$
%&&
'
())
Likelihood of h(Ttruth) given yo: 1
2!" o
exp #h(Ttruth ) # yo( )
2
2" o
2
$
%&&
'
())
Joint likelihood of Ttruth: 1
2!" b
exp #Ttruth # Tb( )
2
2" b
2#h(Ttruth ) # yo( )
2
2" o
2
$
%&&
'
())
Minimizing the cost function maximizes the likelihood of the estimate of truth
![Page 19: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/19.jpg)
From a 3D-Var point of view,we want to find (Ta -Tb) thatminimizes the cost function J.This maximizes the likelihood ofTa~Ttruth given both Tb and yo
2Jmin
=(Ta ! Tb )
2
" b
2+(h(Ta ) ! yo )
2
" o
2
Toy temperature data assimilation, variational approach
Again, we have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
![Page 20: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/20.jpg)
From a 3D-Var point of view,we want to find (Ta -Tb) thatminimizes the cost function J:
So that from
Now
(Ta ! Tb )1
" b
2+H
2
" o
2
#
$%&
'(= (Ta ! Tb )
1
" a
2= H
(yo ! h(Tb ))" o
2
h(Ta ) ! yo = h(Tb ) ! yo + H (Ta ! Tb )
!J / !(Ta" T
b) = 0 we get
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
![Page 21: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/21.jpg)
From a 3D-Var point of view,we want to find (Ta -Tb) thatminimizes the cost function J:
or
So that from
Now
(Ta ! Tb )1
" b
2+H
2
" o
2
#
$%&
'(= H
(yo ! h(Tb ))" o
2
h(Ta ) ! yo = h(Tb ) ! yo + H (Ta ! Tb )
!J / !(Ta" T
b) = 0 we get
w = !b
"2+ H!
o
"2H( )
"1
H!o
"2= !
a
2H!
o
"2
Ta = Tb + w yo ! h(Tb )( ) where now
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
J(Ta ) =(Ta ! Tb )
2
2" b
2+(h(Ta ) ! yo )
2
2" o
2
![Page 22: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/22.jpg)
From a 3D-Var point of view,we want to find (Ta -Tb) thatminimizes the cost function J:
2Jmin
=(Ta ! Tb )
2
" b
2+(h(Ta ) ! yo )
2
" o
2
w = !b
"2+ H!
o
"2H( )
"1
H!o
"2= !
a
2H!
o
"2
where
Toy temperature data assimilation, variational approach
We have a forecast Tb and a radiance obs yo = h(Tt ) + !0
yo ! h(Tb )Innovation:
This variational solution is the same as the one obtained with Kalmanfilter (sequential approach, like Optimal Interpolation):
Ta = Tb + w(yo ! h(Tb )) = Tb + w("0 ! H"b )
with w = !b
2H (!
o
2+!
b
2H
2)"1 (Show the w’s are
the same!)
![Page 23: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/23.jpg)
Typical 6-hour analysis cycle
Forecast phase, followed by Analysis phase
Typical 6-hour analysis cycle
![Page 24: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/24.jpg)
Toy temperature analysis cycle (Kalman Filter)
Forecasting phase, from ti to ti+1: Tb(ti+1) = m T
a(ti)[ ]
So that we can predict the forecast error variance
Forecast error: !b(ti+1) = T
b(ti+1) " T
t(ti+1) =
m Ta(ti)[ ]" m T
t(ti)[ ] + !
m(ti+1) = M!
a(ti) + !
m(ti+1)
! b
2(ti+1) = M
2! a
2(ti ) +Qi; Qi = "m
2(ti+1)
(The forecast error variance comes from the analysis and model errors)
![Page 25: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/25.jpg)
Toy temperature analysis cycle (Kalman Filter)
Forecasting phase, from ti to ti+1: Tb(ti+1) = m T
a(ti)[ ]
So that we can predict the forecast error variance
Now we can compute the optimal weight (KF or Var, whichever form ismore convenient, since they are equivalent):
Forecast error: !b(ti+1) = T
b(ti+1) " T
t(ti+1) =
m Ta(ti)[ ]" m T
t(ti)[ ] + !
m(ti+1) = M!
a(ti) + !
m(ti+1)
! b
2(ti+1) = M
2! a
2(ti ) +Qi; Qi = "m
2(ti+1)
w = !b
2H (!
o
2+ H!
b
2H )
"1= !
b
"2+ H!
o
"2H( )
"1
H!o
"2
(The forecast error variance comes from the analysis and model errors)
![Page 26: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/26.jpg)
Toy temperature analysis cycle (Kalman Filter)
Analysis phase: we use the new observation
Ta (ti+1) = Tb (ti+1) + wi+1 yo(ti+1) ! h Tb (ti+1)( )"# $%
we get
We also need the compute the new analysis error variance:
!a
2(ti+1) =
!o
2!b
2
!o
2+ H
2!b
2
"
#$%
&'i+1
= (1( wi+1H )!
b
2
i+1< !
b
2
i+1
yo(ti+1)
!a
"2= !
b
"2+ H!
o
"2H
now we can advance to the next cycle ti+2, t
i+3,...
compute the new observational increment yo(ti+1) ! h Tb (ti+1)( )
and the new analysis:
from
![Page 27: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/27.jpg)
Summary of toy system equations (for a scalar)
“We use the model to forecast Tb and toupdate the forecast error variance from t
ito ”t
i+1
Tb(ti+1) = m T
a(ti)[ ] !
b
2(ti+1) = M
2 !a
2(ti)"# $% M = !m / !T
Interpretation…
![Page 28: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/28.jpg)
Summary of toy system equations (for a scalar)
Ta = Tb + w yo ! h Tb( )"# $%
“We use the model to forecast Tb and toupdate the forecast error variance from t
ito ”t
i+1
Tb(ti+1) = m T
a(ti)[ ]
At ti+1
“The analysis is obtained by adding to the background theinnovation (difference between the observation and the firstguess) multiplied by the optimal weight:
!b
2(ti+1) = M
2 !a
2(ti)"# $% M = !m / !T
![Page 29: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/29.jpg)
Summary of toy system equations (for a scalar)
Ta = Tb + w yo ! h Tb( )"# $%
“We use the model to forecast Tb and toupdate the forecast error variance from t
ito ”t
i+1
Tb(ti+1) = m T
a(ti)[ ]
At ti+1
“The analysis is obtained by adding to the background theinnovation (difference between the observation and the firstguess) multiplied by the optimal weight:
w = !b
2H (!
o
2+ H!
b
2H )
"1
“The optimal weight is the background error variance dividedby the sum of the observation and the background errorvariance. ensures that the magnitudes and unitsare correct.”
H = !h / !T
!b
2(ti+1) = M
2 !a
2(ti)"# $% M = !m / !T
![Page 30: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/30.jpg)
Summary of toy system equations (cont.)
w = !b
2H (!
o
2+ H!
b
2H )
"1
“The optimal weight is the background error variance dividedby the sum of the observation and the background errorvariance. ensures that the magnitudes and unitsare correct.”
H = !h / !T
Note that the larger the background error variance, thelarger the correction to the first guess.
![Page 31: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/31.jpg)
Summary of toy system equations (cont.)
!a
2=
!o
2!b
2
!o
2+ H
2!b
2
"
#$%
&'= (1( wH )!
b
2
The analysis error variance is given by
“The analysis error variance is reduced from the backgrounderror by a factor (1 - scaled optimal weight)”
![Page 32: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/32.jpg)
Summary of toy system equations (cont.)
!a
2=
!o
2!b
2
!o
2+ H
2!b
2
"
#$%
&'= (1( wH )!
b
2
The analysis error variance is given by
This can also be written as
!a
"2= !
b
"2+!
o
"2H
2( )
“The analysis precision is given by the sum of the backgroundand observation precisions”
“The analysis error variance is reduced from the backgrounderror by a factor (1 - scaled optimal weight)”
![Page 33: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/33.jpg)
Equations for toy and real huge systems
These statements are important because they hold true fordata assimilation systems in very large multidimensionalproblems (e.g., NWP).
Instead of model, analysis and observational scalars, wehave 3-dimensional vectors of sizes of the order of 107-108
![Page 34: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/34.jpg)
Equations for toy and real huge systems
These statements are important because they hold true fordata assimilation systems in very large multidimensionalproblems (e.g., NWP).
We have to replace scalars (obs, forecasts) by vectors
Instead of model, analysis and observational scalars, wehave 3-dimensional vectors of sizes of the order of 107-108
Tb ! xb; Ta ! xa; yo ! yo;
and their error variances by error covariances:
!b
2" B; !
a
2" A; !
o
2" R;
![Page 35: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/35.jpg)
Equations for toy and real huge systems
These statements are important because they hold true fordata assimilation systems in very large multidimensionalproblems (e.g., NWP).
We have to replace scalars (obs, forecasts) by vectors
Instead of model, analysis and observational scalars, wehave 3-dimensional vectors of sizes of the order of 107-108
Tb ! xb; Ta ! xa; yo ! yo;
and their error variances by error covariances:
!b
2" B; !
a
2" A; !
o
2" R;
![Page 36: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/36.jpg)
Interpretation of the NWP system of equations
xa= x
b+K y
o! H x
b( )"# $%
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
At ti+1
”
![Page 37: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/37.jpg)
Interpretation of the NWP system of equations
xa= x
b+K y
o! H x
b( )"# $%
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
At ti+1
“The analysis is obtained by adding to the background theinnovation (difference between the observation and the firstguess) multiplied by the optimal Kalman gain (weight) matrix”
K = BHT(R +HBH
T)!1
”
![Page 38: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/38.jpg)
Interpretation of the NWP system of equations
xa= x
b+K y
o! H x
b( )"# $%
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
At ti+1
“The analysis is obtained by adding to the background theinnovation (difference between the observation and the firstguess) multiplied by the optimal Kalman gain (weight) matrix”
K = BHT(R +HBH
T)!1
“The optimal weight is the background error covariance divided bythe sum of the observation and the background error covariance.
ensures that the magnitudes and units are correct.The larger the background error variance, the larger the correctionto the first guess.”
H = !H / !x
”
![Page 39: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/39.jpg)
Interpretation of the NWP system of equations
xa= x
b+K y
o! H x
b( )"# $%
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
At ti+1
“The analysis is obtained by adding to the background theinnovation (difference between the observation and the firstguess) multiplied by the optimal Kalman gain (weight) matrix”
K = BHT(R +HBH
T)!1
“The optimal weight is the background error covariance divided bythe sum of the observation and the background error covariance.
ensures that the magnitudes and units are correct.The larger the background error variance, the larger the correctionto the first guess.”
H = !H / !x
”
![Page 40: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/40.jpg)
Interpretation of the NWP system of equations
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
”
Forecast phase:
![Page 41: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/41.jpg)
Interpretation of the NWP system of equations
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
”
Forecast phase:
“We use the linear tangent model and its adjoint toforecast B”
B(ti+1) =M A(t
i)[ ]MT
![Page 42: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/42.jpg)
Interpretation of the NWP system of equations
“We use the model to forecast from ti
to ti+1
xb(ti+1) = M x
a(ti)[ ]
”
Forecast phase:
“We use the linear tangent model and its adjoint toforecast B”
B(ti+1) =M A(t
i)[ ]MT
“However, this step is so horrendously expensive that itmakes Kalman Filter completely unfeasible”.
“Ensemble Kalman Filter solves this problem by estimatingB using an ensemble of forecasts.”
![Page 43: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/43.jpg)
Summary of NWP equations (cont.)
A = I !KH( )B
The analysis error covariance is given by
“The analysis covariance is reduced from the backgroundcovariance by a factor (I - scaled optimal gain)”
![Page 44: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/44.jpg)
Summary of NWP equations (cont.)
A = I !KH( )B
The analysis error covariance is given by
This can also be written as
A!1= B
!1+H
TR
!1H
“The analysis precision is given by the sum of the backgroundand observation precisions”
“The analysis covariance is reduced from the backgroundcovariance by a factor (I - scaled optimal gain)”
![Page 45: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/45.jpg)
Summary of NWP equations (cont.)
A = I !KH( )B
The analysis error covariance is given by
This can also be written as
A!1= B
!1+H
TR
!1H
“The analysis precision is given by the sum of the backgroundand observation precisions”
“The analysis covariance is reduced from the backgroundcovariance by a factor (I - scaled optimal gain)”
K = BHT(R +HBH
T)!1= (B
!1+H
TR
!1H)
!1H
TR
!1
“The variational approach and the sequential approach aresolving the same problem, with the same K, but only KF (orEnKF) provide an estimate of the analysis error covariance”
![Page 46: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/46.jpg)
3D-Var3D-Var
J = min
1
2[(xa
! xb )T
B!1(xa
! xb ) + (Hx
a! y)T
R!1(Hx
a! y)]
Distance to forecast Distance to observations
4D-Var4D-Var
J = min1
2[(x
0! x
0
b )TB
!1(x0! x
0
b ) + (Hxi! y
i)T
Ri
!1(Hxi! y
i)
i=1
s
" ]
at the analysis time
Distance to background at theinitial time
Distance to observations in atime window interval t0-t1
Control variable x(t0) Analysis x(t
1) = M[x(t
0)]
It seems like a simple change, but it is not! (e.g., adjoint)It seems like a simple change, but it is not! (e.g., adjoint)What is B? It should be tunedWhat is B? It should be tuned……
![Page 47: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/47.jpg)
Ensemble Transform Kalman FilterEnsemble Transform Kalman Filter(EnKF)(EnKF)
Forecast step:
Analysis step:
The new analysis error covariance in the ensemble space is (Hunt et al.2007)
And the new ensemble perturbations are given by (transform)
x
n,k
b= M
nx
n!1,k
a( )Bn=
1
K !1X
n
bX
n
bT, where X
n
b= x
n,1
b ! xn
b;...,x
n,K
b ! xn
b"# $%
x
n
a= x
n
b+ K
n(y
n! Hx
n
b ); Kn= B
nH
T (R + HBnH
T )!1
!An= K !1( )I + (HX
n
b)TR
!1(HX
n
b)"# $%
!1
X
n
a= X
n
bK !1( ) !A
n"# $%
1/2
![Page 48: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/48.jpg)
Comparison of 4-D Var and LETKF at JMAT. Miyoshi and Y. Sato
• 4D-Var and EnKF are the two advanced, feasible methods• There will be a workshop on them in Buenos Aires (Nov’08)!!!
• In Ensemble Kalman Filter the background error covarianceB is approximated and advanced in time with an ensemble ofK forecasts. In the subspace of the ensemble, B=I so thatmatrix inversions are efficient.
• So far, comparisons show EnKF is slightly better than 3D-Var, but there has not been enough time to develop tunings
• At JMA, Takemasa Miyoshi has been performingcomparisons of the Local Ensemble Transform Kalman Filter(Hunt et al., 2007) with their operational 4D-Var
• Comparisons are made for August 2004
![Page 49: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/49.jpg)
Comparison of 4D-Var and LETKF at JMAT. Miyoshi and Y. Sato
N.H.
S.H.
Tropics
ACRMS error against analysis
Bias
![Page 50: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/50.jpg)
Comparison of LETKF and 4D-Var at JMAT. Miyoshi and Y. Sato
N.H.
S.H.
Tropics
VerifyingagainstRawinsondes
RMSerror
Bias
![Page 51: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/51.jpg)
Comparison of 4-D Var and LETKF at JMA18th typhoon in 2004, IC 12Z 8 August 2004
T. Miyoshi and Y. Sato
operational LETKF
![Page 52: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/52.jpg)
Comparison of 4-D Var and LETKF at JMARMS error statistics for all typhoons in August 2004
T. Miyoshi and Y. Sato
Operational 4D-Var LETKF
![Page 53: Introduction to data assimilation and least squares methods4dvarenkf.cima.fcen.uba.ar/course/download/Kalnay...Toy temperature data assimilation, variational approach We have a forecast](https://reader035.fdocuments.in/reader035/viewer/2022062607/6044935305ae353c4408f4ac/html5/thumbnails/53.jpg)
Summary• Data assimilation methods have contributed much to the
improvements in NWP.• A toy example is easy to understand, and the equations
are the same for a realistic system• Kalman Filter (too costly) and 4D-Var (complicated) solve
the same problem (if model is linear and we use longassimilation windows)
• Ensemble Kalman Filter is feasible and simple• It is starting to catch up with operational 4D-Var• EnKF can also estimate observational errors online• Important problems: estimate and correct model errors &
obs. errors, optimal obs. types and locations, tuningadditive/multiplicative inflation, parameters estimation,…– Tellus: 4D-Var or EnKF? Tellus 2007– Papers posted in “Weather Chaos UMD”– Workshop in Buenos Aires Nov ’08