Transcript KentBias_SAMOS_v2

Bias Correction of
In Situ Observations
Elizabeth Kent, David Berry, Peter Taylor,
Margaret Yelland and Ben Moat
National Oceanography Centre, Southampton, UK
(with help from Peter Challenor, Dick Reynolds
and Tom Smith)
Outline
o Sources of bias
o Identification of bias
• Using high-quality data (examples: OWS air temperature,
RV winds)
• Using biased data (example: VOS SST)
• Using comparison standards
o Model output (examples: VSOP-NA, VOSClim and MetDB)
o Satellite retrievals (example: scatterometer winds)
• Modelling (example: airflow distortion)
• 3-way intercomparisons (example sea surface heights)
o Metadata
o Conclusions
Some sources of bias
o Winds
• Airflow distortion
• Height adjustment
• Anemometer calibration
• Visual/anemometer inhomogeneity
o SST
• Variations in measurement depth (definition of SST)
• Sensor calibration/drift
• Heating/cooling of local environment
• Mixing of surface layer by platform
Some sources of bias …. continued
o Air Temperature
• Radiative heating
• Height adjustment
• Sensor calibration/drift
• Sensor ventilation
o Humidity
• As air temperature plus drying of wet bulb
• Salt contamination
o Pressure
• Sensor calibration/drift
• Wind effects
• Sealed bridge (air conditioning).
Identification of bias: Using high quality data
o Air temperature data from
OWS L used to derived
correction for radiative
heating.
o Comparison of wellexposed and poorlyexposed sensors used to
quantify biases.
o Biases modelled with
analytical model of heat
budget (Berry et al., 2004:
An analytical model of
heating errors in marine air
temperatures from ships J.
Atmos. Ocean. Tech.,
21,1198 - 1215.)
Identification of bias: Using high quality data
o Comparison of Research
Vessel winds and pressures
with reanalysis output
reveals bias in NCEP1.
o NCEP1 winds are too low,
except at very low wind
speeds and the bias
increases with increasing
wind speed.
o Pressure gradients are
underestimated in NCEP1
o See Smith et al., 2001:
Quantifying Uncertainties in
NCEP Reanalyses Using
High-Quality Research
Vessel Observations, J.
Climate., 14,4062 - 4072.
Identification of bias: Using biased data
o We do not need unbiased
data to reveal biases, but
we must understand the
error structure of the
data.
o A comparison of bucket
and engine intake SST
revealed biases in both
sources of data.
o Kent, and Kaplan, 2006:
Toward Estimating
Climatic Trends in SST,
Part 3: Systematic
Biases. J. Atmos. Ocean.
Tech., 23, 487-500.
Identification of bias: Using biased data
o
Comparison of sonic anemometers
on either side of the RRS
Discovery foremast.
o
The effects of flow distortion by the
foremast platform and extension
are seen in the ratio of the two
sensors.
o
We should be able to choose the
least-biased sensor as a function of
relative wind direction.
o
Yelland et al., 2002: CFD model
estimates of the airflow over
research ships and the impact on
momentum flux measurements. J.
Atmos. Oceanic Tech.,19,14771499.
Identification of bias: Comparison
standards
o It it not necessary that
the standard used is bias
free.
o Useful to compare
observations made by
different methods.
o Examples using model
output:
• VSOP-NA
• VOSClim
• Met Office GTS
Monitoring reports
Identification of bias: Intercomparison
o Need to remove large scale
biases in AVHRR satellite SST
using in situ data.
o But buoy and ship SST have
relative bias too.
o Biases are estimated by using
Empirical Orthogonal
Teleconnections (EOT).
o Plots show the estimated bias
in Pathfinder AVHRR SST for
2003-2004, before (top) and
after (bottom) bias correction
using these techniques.
o Smith and Reynolds, 2005,
Journal of Climate, 18, 20212036.
Identification of bias: Comparison Standards
o Satellite retrievals can also
be used to compare
distributed in situ reports.
o The number of colocations is often small.
o Again need to consider
both random and
systematic errors in both
data sources.
o Kent, Taylor and
Challenor, 1998: A
Comparison of Ship and
Scatterometer-Derived
Wind Speed Data.
Int.J.Remote Sensing,
19(17), 3361-3381.
Identification of Bias: Modelling airflow
Identification of Bias: Laboratory measurements
o Buckets used to sample
the sea water for SST
measurements are
insulated to reduce heat
exchange with the
atmosphere.
o However analysis
suggested that heat
exchange could be
detected.
o Laboratory measurements
of the temperature of warm
water in the bucket over
time showed that this was
likely.
Identification of bias: Triple co-locations
o Given 3 independent
estimates of the same
quantity (here sea surface
height from a model and 2
satellites) can calculate the
errors in each.
o If the errors are not
independent (e.g. errors in
tide correction to satellite
estimates) this can be
accounted for.
o Cannot exclude the possibility
of bias common to all.
o Tokmakian, and Challenor,
2000, Ocean Modelling., 1,
39-52 or Caires and Sterl,
Geophys. Res., 2003, C108,
3, 3098,
doi:10.1029/2002JC001491.
Metadata
o Metadata is necessary to analyse bias and apply
corrections
• Measurement heights
• Measurement methods
• Instrument and calibration information
• Instrument siting
o Metadata availability is patchy
• VOS metadata is collected in Pub. 47 (1955-present)
• Historical ODAS metadata can be hard to get hold of
• Research Vessel metadata ranges from excellent to nonexistent
o Pragmatic decisions - if you ask for too much you can
end up with nothing.
o Proxy and implied metadata can be useful.
Metadata: Deduction from the data
See: Berry and Kent,
2005: The Effect of
Instrument Exposure
on Marine Air
Temperatures: An
Assessment Using
VOSClim Data, Int.
J. Climatol., 25,
1007-1022.
Metadata: Proxies
July 1980-1997: Air Temperature Differences from
Local Mean Normalised by Local Standard Deviation
Conclusions
o A range of methods is available to quantify bias.
o Need to consider biases in all data sources.
o Random errors can appear systematic if not
handled correctly.
o Metadata are vital in the quantification and
correction of bias.
o Sometimes we may be able to deduce metadata
from the characteristics of the data themselves.
o Bias correction can rehabilitate data for a wide
range of applications.