Introduction to data assimilation
and least squares methods
Eugenia Kalnay
and many friends
University of Maryland
October 2008 (part 1)
The observing system a few years ago
Before
1979
only
raobs
Now we have even more satellite data
Typical distribution of the observing systems in a 6 hour period:
Typical
of observations
in +/-times
3hours
a realdistribution
mess: different
units, locations,
Model grid points (uniformly distributed) and observations
(randomly distributed). For the grid point i only observations
within a radius of influence may be considered
i
k
Intro. to data assimilation: toy example 1
We want to measure the temperature in this
room, and we have two thermometers that
measure with errors:
T =T +
1
T2 = Tt + 2
We assume that the errors are unbiased:
1 = 2 = 0
that we know their variances 12 = 12 22 = 22
and the errors of the two thermometers are
uncorrelated: 1 2 = 0
The question is: how can we estimate the true
temperature optimally? We call this optimal
estimate the analysis of the temperature
Intro. to data assimilation: toy example 1
We try to estimate the analysis from a linear
combination of the observations:
Ta = a1T1 + a2T2
and assume that the analysis errors are unbiased:
Ta = Tt
a1 + a2 = 1
This implies that
Intro. to data assimilation: toy example 1
We try to estimate the analysis from a linear
combination of the observations:
Ta = a1T1 + a2T2
and assume that the analysis errors are unbiased:
Ta = Tt
a1 + a2 = 1
This implies that
Ta will be the best estimate of if the coefficients a1 , a2
are chosen to minimize the mean squared error of Ta:
a2 = (Ta Tt )2 = [a1 (T1 Tt ) + (1 a1 )(T2 Tt )]2
Intro. to data assimilation: toy example 1
Replacing
a2 = 1 a1
2
the minimization of a with respect to
a1 gives
a2 = (Ta Tt )2 = [a1 (T1 Tt ) + (1 a1 )(T2 Tt )]2
Intro. to data assimilation: toy example 1
Replacing
a2 = 1 a1
the minimization of
a2
with respect to
a1
gives
a2 = (Ta Tt )2 = [a1 (T1 Tt ) + (1 a1 )(T2 Tt )]2
a2
= 0 ==>
a1
or
22
a1 = 2
1 + 22
1 / 12
a1 =
1 / 12 + 1 / 22
12
a2 = 2
1 + 22
1 / 22
a2 =
1 / 12 + 1 / 22
The first formula says that the weight of obs 1 is given by the variance
of obs 2 divided by the total error.
The second formula says that the weights of the observations are
proportional to the "precision" or accuracy of the measurements
(defined as the inverse of the variances of the observational errors).
Intro. to data assim: toy example 1 summary
Two measurements and an optimal linear combination (analysis):
2
Optimal
coefficients
(min
)
a
T = aT +a T
a
1 1
a1 = 2
1 + 22
2
2
2 2
a2 = 2
1 + 22
2
1
or
1 / 12
a1 =
1 / 12 + 1 / 22
2 2
2
12
=
Replacing, we get a
12 + 22
or
1 / 22
a2 =
1 / 12 + 1 / 22
1
1
1
= 2+ 2
2
a
1 2
Intro. to data assim: toy example 1 summary
Two measurements and an optimal linear combination (analysis):
2
Optimal
coefficients
(min
)
a
T = aT +a T
a
1 1
a1 = 2
1 + 22
2
2
2 2
a2 = 2
1 + 22
2
1
or
1 / 12
a1 =
1 / 12 + 1 / 22
2 2
2
12
=
Replacing, we get a
12 + 22
or
1 / 22
a2 =
1 / 12 + 1 / 22
1
1
1
= 2+ 2
2
a
1 2
Intro. to data assim: toy example 1 summary
Two measurements and an optimal linear combination (analysis):
2
Optimal
coefficients
(min
)
a
T = aT +a T
a
1 1
a1 = 2
1 + 22
2
2
2 2
a2 = 2
1 + 22
2
1
or
1 / 12
a1 =
1 / 12 + 1 / 22
2 2
2
12
=
Replacing, we get a
12 + 22
or
1 / 22
a2 =
1 / 12 + 1 / 22
1
1
1
= 2+ 2
2
a
1 2
Suppose that T1=Tb (forecast) and T2=To (observation). Then
Ta = a1Tb + a2To = Tb + a2 (To Tb )
Intro. to data assim: toy example 1 summary
Two measurements and an optimal linear combination (analysis):
2
Optimal
coefficients
(min
)
a
T = aT +a T
a
1 1
a1 = 2
1 + 22
2
2
2 2
a2 = 2
1 + 22
2
1
or
1 / 12
a1 =
1 / 12 + 1 / 22
2 2
2
12
=
Replacing, we get a
12 + 22
or
1 / 22
a2 =
1 / 12 + 1 / 22
1
1
1
= 2+ 2
2
a
1 2
Suppose that T1=Tb (forecast) and T2=To (observation). Then
Ta = a1Tb + a2To = Tb + a2 (To Tb )
or
Ta = Tb + 2
(To Tb )
2
b + o
2
b
This is the form that is always used in analyses
Intro. to data assim: toy example 1 summary
A forecast and an observation optimally combined (analysis):
Ta = Tb + 2
(To Tb )
2
b + o
2
b
with
1
1
1
= 2+ 2
2
a b o
If the statistics of the errors are exact, and if the coefficients
are optimal, then the "precision" of the analysis (defined as
the inverse of the variance) is the sum of the precisions of
the measurements.
Now we are going to see a second toy example of data
assimilation including remote sensing.
The importance of these toy examples is that the equations
are identical to those obtained with big models and many obs.
Intro. to remote sensing and data
assimilation: toy example 2
Assume we have an object, a stone in space
We want to estimate its temperature T (oK) accurately but we
measure the radiance y (W/m2) that it emits. We have an obs.
model, e.g.:
y = h(T ) T 4
Intro. to remote sensing and data
assimilation: toy example 2
Assume we have an object, a stone in space
We want to estimate its temperature T (oK) accurately but we
measure the radiance y (W/m2) that it emits. We have an obs.
y = h(T ) T 4
model, e.g.:
We also have a forecast model for the temperature
T (ti +1 ) = m [T (ti )];
e.g., T (ti +1 ) = T (ti ) + t [ SW heating+LW cooling ]
Intro. to remote sensing and data
assimilation: toy example 2
Assume we have an object, a stone in space
We want to estimate its temperature T (oK) accurately but we
measure the radiance y (W/m2) that it emits. We have an obs.
model, e.g.:
y = h(T ) T 4
We also have a forecast model for the temperature
T (ti +1 ) = m [T (ti )];
e.g., T (ti +1 ) = T (ti ) + t [ SW heating+LW cooling ]
We will derive the data assim eqs (KF and Var) for this toy
system (easy to understand!)
Intro. to remote sensing and data
assimilation: toy example 2
Assume we have an object, a stone in space
We want to estimate its temperature T (oK) accurately but we
measure the radiance y (W/m2) that it emits. We have an obs.
model, e.g.:
y = h(T ) T 4
We also have a forecast model for the temperature
T (ti +1 ) = m [T (ti )];
e.g., T (ti +1 ) = T (ti ) + t [ SW heating+LW cooling ]
We will derive the data assim eqs (OI/KF and Var) for this toy
system (easy to understand!)
Will compare the toy and the real huge vector/matrix
equations: they are exactly the same!
Toy temperature data assimilation, measure radiance
We have a forecast Tb (prior) and a radiance obs yo = h(Tt ) + 0
The new information (or innovation) is the
observational increment:
yo h(Tb )
Toy temperature data assimilation, measure radiance
We have a forecast Tb (prior) and a radiance obs yo = h(Tt ) + 0
The new information (or innovation) is the
observational increment:
yo h(Tb )
The final formula is the same:
Ta = Tb + w(yo h(Tb ))
with the optimal weight
w = H ( + H H )
2
b
2
o
2
b
1
Toy temperature data assimilation, measure radiance
We have a forecast Tb (prior) and a radiance obs yo = h(Tt ) + 0
The new information (or innovation) is the
observational increment:
yo h(Tb )
We assume that the obs. and model errors are Gaussian
The innovation can be written in terms of errors:
yo h(Tb ) = h(Tt ) + 0 h(Tb ) = 0 + h(Tt ) h(Tb ) = 0 H b
where H = h / T includes changes of units
and observation model nonlinearity, e.g., h(T ) = T 4
Toy temperature data assimilation, measure radiance
We have a forecast Tb and a radiance obs yo = h(Tt ) + 0
yo h(Tb ) = 0 H b
Toy temperature data assimilation, measure radiance
We have a forecast Tb and a radiance obs yo = h(Tt ) + 0
yo h(Tb ) = 0 H b
From an OI/KF (sequential) point of view:
Ta = Tb + w(yo h(Tb )) = Tb + w( 0 H b )
or
a = b + w( 0 H b )
Now, the analysis error variance (over many cases) is
a2 = a2
Toy temperature data assimilation, measure radiance
We have a forecast Tb and a radiance obs yo = h(Tt ) + 0
yo h(Tb ) = 0 H b
From an OI/KF (sequential) point of view:
Ta = Tb + w(yo h(Tb )) = Tb + w( 0 H b )
a = b + w( 0 H b )
or
In OI/KF we choose w to minimize the analysis error:
We compute
a2 = a2
= + w ( + H H ) 2w H
2
a
2
b
2
o
2
b
assuming that b , 0 are uncorrelated
2
b
Toy temperature data assimilation, measure radiance
We have a forecast Tb and a radiance obs yo = h(Tt ) + 0
yo h(Tb ) = 0 H b
From an OI/KF (sequential) point of view:
Ta = Tb + w(yo h(Tb )) = Tb + w( 0 H b )
a = b + w( 0 H b )
or
In OI/KF we choose w to minimize the analysis error:
a2 = a2
a2 = b2 + w 2 ( o2 + H b2 H ) 2w b2 H
From
=0
w
2
a
we obtain
w = b2 H ( o2 + H b2 H )1
Toy temperature data assimilation, measure radiance
Repeat: from an OI/KF point of view the analysis (posterior) is:
Ta = Tb + w(yo h(Tb )) = Tb + w( 0 H b )
with
w = b2 H ( o2 + b2 H 2 )1
Note that the scaled weight
If
If
wH
is between 0 and 1
>> H
Ta Tb
<< H
Ta wyo
2
o
2
o
2
b
2
b
The analysis interpolates between the background and the
observation, giving more weights to smaller error variances.
Toy temperature data assimilation, measure radiance
Repeat: from an OI/KF point of view the analysis (posterior) is:
Ta = Tb + w(yo h(Tb )) = Tb + w( 0 H b )
with
w = b2 H ( o2 + b2 H 2 )1
Subtracting
Tt
from both sides we obtain
a = b + w( 0 H b )
Squaring the analysis error and averaging over many cases, we obtain
= (1 wH )
2
a
2
b
which can also be written as
1 1 H2
= 2 + 2
2
a b o
Toy temperature data assimilation, measure radiance
Summary for OI/KF (sequential):
Ta = Tb + w(yo h(Tb ))
with
analysis
w = H ( + H )
2
b
2
o
2
b
2 1
optimal weight
The analysis error is computed from
a2 = (1 wH ) b2
which can also be written as
1 1 H2
= 2 + 2
2
a b o
analysis precision=
forecast precision + observation precision
