Transcript Here

Barcelona, June 2016
www.bsc.es
Assessment of Antarctic sea ice
thickness in the ORA-IP
Contribution to Evaluation of Ocean
Syntheses action in the Polar Regions
François Massonnet
Antarctic sea ice thickness at a glance
•
Among all essential climate variables, Antarctic sea ice thickness (AA-SIT) is
probably one with poorest coverage.
•
AA-SIT is the product of a subtle balance between large atmospheric and
oceanic heat fluxes and is almost entirely seasonal. The Antarctic pack is
essentially divergent and subject to very powerful negative feedbacks; ice is thin
(~1 m on avg).
•
Recent satellite technology has allowed SIT retrievals but large uncertainties
remain due to substantial snow load.
•
The proportion of deformed, and therefore locally thick Antarctic sea ice, is
probably more deformed than previously thought (Williams et al., Nat. Geosci.
2015)
It is important to assess to what extent current reanalyses
simulate realistically (or not) Antarctic sea ice thickness. If they
do, these reanalyses can help to estimate the mass balance of
sea ice (and perhaps more) in the Southern Hemisphere.
1
ASPeCt thickness: heterogeneous
and subject to thin bias
•
ASPeCt: “Antarctic Sea Ice Processes and Climate”
•
•
•
•
•
A unique data set: 1981-2005, all sectors of Antarctica
81 voyages + 2 helicopter flights nearly everywhere
Over 23,000 individual measurements of sea ice thickness
Level ice thickness, weighted for open water, is considered here
(equal to sea ice volume divided by area of region of sampling,
including open water, excluding deformed ice).
Data is subject to large uncertainties (and error statistics are not
provided)
•
•
•
Representativity error: data were collected during localized
expeditions
Systematic bias: towards thin ice (ice-breaker-based measurements)
Measurement error: “The thickness is estimated on the level parts of
floes when they are broken and turned sideways along the hull of the
ship” (Worby et al., JGR, 2008). In addition the snow/ice interface is
difficult to localize visually from a ship.
A careful evaluation process has to be proposed
More info on ASPeCt: Worby et al., JGR, 2008
2
Rebinning ASPeCt data
All 23,391 daily ASPeCt measurements (1981-2005) were first binned
into 1°x1° grid boxes, for each month of each year. The mean, min,
max, and number of data for each bin were retained. The number of
daily observations follows a power law:
That is, there are less than three measurements available for 50% of the
cases, meaning that any obs-model comparison must account for possible
sampling issues.
By comparison, ORA-IP reanalyses provide the monthly-mean sea ice
thickness over 1°x1° boxes.
3
Are the ORA-IP compatible
with the ASPeCt data?
For each grid box and each month, we want to verify whether the collected ASPeCt
samples are compatible (statistically speaking) with the mean provided by some
ORA-IP reanalysis.
When only one, two or three observations are available (~50% of the cases), no
evaluation is conducted: sample size is just too small and nothing robust can be
concluded since nothing is known about the variance of the SIT distribution.
When four or more observations, the ORA-IP is deemed successful it its mean
value lies within the range of ASPeCt data. Assuming symmetrical PDF, the
probability for Type-I error is 12.5% (1 / 24 – 1)
Note: the test has limitations!
ORA-IP mean value,
assuming it is correct
PDF of true thickness
ORA-IP is
compatible
with ASPeCT
The test is
positive
BUT
The test is
positive
Five ASPeCt measurements
ORA-IP is
compatible
with ASPeCt
Or, stated differently: a negative test indicates mismatch
but a positive test does not necessarily indicate a match.
4
Are the ORA-IP compatible
with the ASPeCt data?
For each month of each year, and each 1°x1° grid box,
- If less than 3 observations are available during that month and at that location,
no assessment is conducted
- Else, « 1 » is assigned to that grid box for that month if the reanalyzed SIT is
within the ASPeCt range (« 0 » otherwise)
- The number of « succeed » and « fail » is cumulated Reanalysis value: succeeds
Example July 1995
(within range)
1°
1°
1
ASPeCt
measurements
Ice thickness
0
Reanalysis value: fails
(outside range)
Reanalysis value: undefined
(not enough ASPeCt data)
Ice thickness
ASPeCt
measurements
ASPeCt
measurements
Ice thickness
5
Compatibility index
# successes
Compatibility index =
# successes + # fails
Here: CI = 3 / (3 + 4) = 0.43
6
Ensemble of reanalyses considered
Data were downloaded from the ORA-IP FTP: ftp://ftp.icdc.zmaw.de/ora_ip/
Thanks to those who organized this server and the datasets!
The evaluation is conducted on 1993-2005 (longest common period
between reanalyses and ASPeCt data).
Label
Institution
Compatibility
index
Mean abs
error (cm)
GloSea5
UK Met Office 0.44
13
GECCO2
U. Hamburg
0.28
22
ECDA
GFDL
0.48
13
GLORYSv1
MERCATOR
0.42
13
GLORYSv3
MERCATOR
0.43
13
C-GLORS
CMCC
0.49
13
ECCO
NASA
0.44
12
Given the heterogeneity of the ASPeCt data set, it is difficult to make absolute
statements. Nevertheless, none of the reanalysis simulates more than 50% of
the time a SIT value that lies within the ASPeCt range. The ORA-IPs cannot be
considered consistent with that dataset (Type-I probability of error: 12.5%)
7
www.bsc.es
Thank you!
For further information please contact
[email protected]