Monitoring Climate Change using
Satellites: Lessons from MSU
Peter Thorne, Simon Tett
Hadley Centre, Met Office, Exeter, UK
UAH data from John Christy
Residual uncertainty work in collaboration with John Christy, Roy
Spencer and David Parker
1
What is the problem?
„
„
2
Forming a homogeneous series from several different
satellites
Corrections are required for:
– Orbit decay -- satellite gets closer to Earth
» Only needed for LT retrieval and v. small uncert.
– Diurnal drift -- satellites drift aliasing in the diurnal
cycle
– Instrument temperature.
» Conversion into brightness temperature has nonlinear dependence on the satellite temperature.
– Other intra-satellite bias.
» Any remaining biases removed.
– Inter-satellite biases
Dataset trend uncertainty
„
Two sources:
– Structural uncertainty
» the uncertainty introduced by the method chosen to go
from raw radiances to a “homogeneous dataset”
– Residual uncertainty
» Uncertainty inherent in the method in the presence of
finite data.
3
Three MSU datasets
MT
trend
Intra-satellite
changes
0.02K
Raw
radiances
UAH
Grey box
0.10K
RSS
Inter-satellite
changes
4
Decisions made
will always
involve a degree
of subjectivity in
absence of agreed
transfer standards
0.24K
V&G
What is the true structural
uncertainty?
P(x)
OR
???
Red is the PDF of best-guess global-mean trends for an
infinite number of physically realistic treatments, green stars
published estimates.
Which (L or R) is correct is important!
Are the datasets consistent?
„
The respective published estimates with 2 sigma (residual
only) uncertainty estimates are:
– UAH: 0.02 +/-0.05 K / decade
– RSS: 0.10 +/-0.02 K / decade
– V & G: 0.24 +/-0.02 K / decade
„
Implies either:
1. some (all?) are physically implausible methods or
2. that structural uncertainty is the major source of
uncertainty (error!) and that this implicitly needs to be
taken into account:
– How? We are grossly under-sampling the structural uncertainty
phase space.
6
Residual dataset uncertainty
„
How were these uncertainty estimates derived?
– Could they simply be under-estimated?
» Might a more realistic set of residual uncertainty
estimates obviate the need to consider structural
uncertainty because it is in fact unimportant?
„
7
Concentrate on UAH as it has had most analysis
applied to it, but similar principles will pertain to
the other datasets.
Internally and Externally derived
error estimates
„
Attempt to produce “internal” error estimates
– “Model” the various components of treatment error to estimate
total error.
» Need to get major error sources and be right about “model”
» Allows computation of any desired quantity.
» Independent
„
Alternatively produce “external” error estimates
– by comparison with radiosondes
» Need enough radiosonde data
» Need to assume error distribution (as sondes are sparse)
» Radiosondes contain errors!
8
Inter-Satellite bias
„
Chosen as one example for internally derived
estimate.
„
Uncertainty in bias is normal expression for
standard deviation (σ/√N) where N is the
estimated dof.
Estimate 1-σ error from 90-day averages (indep.
data)
„
„
9
Biases are cumulative.
LT Inter-Satellite differences
Drift in NOAA-12
10
Bias Uncertainties
Product
11
Tropics Global
Pre NOAA-12 0.034
0.031
LT NOAA-12 0.052
0.037
post NOAA-12 0.021
0.015
MT
0.024
0.018
LS
0.071
0.063
Externally derived estimates
Ten U.S. VIZ sites RMT vs. TMT 0-30N
1.5
1.0
0.5
0.0
°C
-0.5
Sonde change
RMT
-1.0
TMT
RMT-TMT
-1.5
-2.0
0.0 °C
-2.5
-0.5
Trend in Difference (unadjusted) = -0.05 °C/decade
-3.0
1979
1981
1983
1985
1987
1989
1991
1993
1995
1997
1999
Fig. 5 Christy et al. 2003
12
2001
Results
„
„
MSU/AMSU Atmospheric Weighting Functions .
Three products (LT,
MT & LS)
0
TLS Lower Stratosphere .
200
Two regions:
TMT Mid-Troposphere .
– tropics (+/- 20o)
– Global
400
hPa
600
TLT Lower Troposphere .
800
1000
0
0.01
0.02
Weight per 5 hPa Layer
0.03
0.04
0.05
J.R. Christy and R.W. Spencer
University of Alabama in Huntsville .
0.06
Errors in the trend
14
Residual errors
15
„
Analysis of UAH shows that residual error
estimates are not likely to be (at least grossly)
underestimated.
„
The two remaining MSU datasets need a thorough
error analysis and this needs to be published.
Lessons?
16
„
Critically important to place robust error estimates.
„
But, structural uncertainty may be the major source
of error: if so this is a big challenge!
„
Having three independently produced estimates
permits an in-depth analysis which is unlikely to
be possible for other satellite datasets and will
undoubtedly provide valuable information.
Just a satellite problem?
17
„
What is happening to tropospheric temperatures
fundamentally affects our understanding of climate
change.
„
Depending upon which MSU series you choose the
answer changes absolutely.
„
We desperately need a clear-cut and objectively
based answer as to what the true trend is with error
estimates!
„
Needs expert input from the satellite, climate,
reanalysis, and observational communities.