AMIP II Sea Surface Temperature and Sea Ice Concentration Observations

Mike Fiorino

Program for Climate Model Diagnosis and Intercomparison 

Lawrence Livermore National Laboratory 

Livermore, CA 94950

19 February, 1997
15 September, 2000

Disclaimer    This document hopes to promote feedback between PCMDI, data providers and the AMIP community. While the observations for the AMIP II boundary conditions have been finalized your comments and suggestions are encouraged. Please respond by sending email to Mike at fiorino@llnl.gov.  

Introduction

The specification of sea surface temperature (SST) and sea ice concentration (SIC) is the most basic and perhaps the most important experimental condition for AMIP II. The AMIP I boundary conditions (BCS) suffered from two basic deficiencies: 1) large and unrealistic changes in sea ice coverage, particularly in the Antarctic; and 2) undefined points that led to substantial and unintended differences in the actual SST used in the 30+ models, especially around the coastlines. The first problem is data related while the second is primarily technical.

To overcome these observation deficiencies, and to improve the quality of the boundary forcing, we have created AMIP II BCS observations (BCSOBS) that are:

• Consistent with the BCSOBS used in the ECMWF and NCEP reanalysis
• Most current and best available
• Appropriate for climate model applications.
• Include the land and glacier/ice-shelf points masks used in the development
• Global grids with data defined at all points

Consistency with reanalysis is desirable as reanalysis will provide the bulk of AMIP II validation data. However, a considerable effort has gone into improving the BCS data sets since the reanalysis BCS were set two years ago. This is particularly true for SIC. While consistency with reanalysis is desirable, we wanted BCSOBS best representative of the relatively observationally rich 17-year AMIP II period as opposed to BCSOBS designed for multi-decadal GCM integrations such as the Climate of the 20^th Century (C20C) experiment.

Another deficiency in the AMIP I BCS was a smoothed and reduced amplitude of the SST annual cycle caused by interpolation of the monthly means (assumed to be valid at the mid point of the month) to daily values. That is, the monthly mean of the daily-interpolated values was different than the original observed input monthly mean. Taylor et al (1996) have developed a procedure for AMIP II to adjust the observations so that the monthly mean of the daily-interpolated BCS is the same as the observed monthly mean (BCSOBS). Thus, the SST and SIC BCS for AMIP II will not be the same as the observations. This paper focuses solely on the observations.

To avoid technical differences in BCS among the AMIP II participants, the BCS have valid data at all points. The method for filling undefined points over land/glaciers is the same algorithm as used in the NCEP reanalysis for setting SST over land and produces smoothly varying fields across the land/sea interface.

The BCS in the NCEP/NCAR reanalysis and ECMWF reanalyses were linearly interpolated to daily values to simplify their use in the assimilation process. When appropriate, we have formed monthly means using these daily values. The source data for these "daily" values, however, come from monthly mean, weekly or daily analyses. The change in frequency of the source data sets was the primary reason for using monthly means instead of daily or weekly BCSOBS in AMIP II.

Source Data

There are three basic sources of data for the AMIP II BCSOBS:

1. satellite-sensed sea ice concentration from A. Nomura (JMA/ECMWF) [Nomura, 1995 #32] and R. Grumbine (NCEP) [Grumbine, 1996 #34] - passive microwave radiation converted to sea ice concentration using the NASA-Team algorithm [Cavalieri, 1992 #92]
2. OISST from NCEP (Optimal Interpolation SST) [Reynolds, 1994 #17] - bias corrected and reanalyzed.
3. GISST2.2a (Global Ice and SST) from the Hadley Centre of the U. K. Meteorological Office [Rayner, 1995 #33] - A new, EOF-reconstructed SST used for the C20C integrations. GISST2.2a is the corrected version of GISST2.2. In the original GISST2.2, the SST in the sea ice margin was in error.

The daily OISST and SIC data were kindly provided by NCEP and are identical to that used in both reanalyses. The GISST2.2a data were provided courtesy of the Hadley Centre.

The properties of the source data for the AMIP II BCSOBS are illustrated using a time line and in the following table.
 

Field	Source	Period	Frequency	Period	Resolution
SST	GISST 2.2a (UKMO)  OISST (NCEP)	7812-8111  8112-present	monthly  weekly	7812-9512	1x1deg
Sea Ice	Nomura (ECMWF)  Grumbine (NCEP)  Grumbine (NCEP)	7812-9111  9111-9510  9511-present	weekly  daily  daily	7812-9511  9512-present	1x1 deg  1x1 deg  0.5x0.5 deg

We first note that the observations have three different frequencies and two horizontal grid resolutions. Furthermore, while most grids had a resolution of 1x1 deg, some had different origins. Thus, a considerable amount of data mechanics was necessary to rectify the data to the common grid (e.g., shifting of the cyclically continuous in longitude grids to the prime meridian).

The most important properties are summarized as:

1. Different sources for both SST and SIC
2. Different times when the source data change
3. Different sea ice data used in the two SST analyses

Preprocessing of the SST Source Data

Both the GISST2.2a and OISST data sets contained internal sea ice masks (SST ~ -1.8 C) derived from different sources. Furthermore, these internal masks differed from that in the Nomura/Grumbine SIC. Thus, the SST and SIC used in reanalysis could conflict. For example, the OISST might have a point which is sea ice (-1.8 C) but the SIC data has the point as open ocean. Conversely, the SIC might define a point as being ice, but the SST would be warmer than the sea ice temperature.

The problem is illustrated for the daily analyses of 1 January, 1996. The white boxes show where the original SIC and SST data were defined as sea ice using 0.55 as the threshold value for the mask. The blue boxes show where the SIC analysis defines open ocean but the OISST is sea ice, and the red boxes where the OISST is open ocean and the SIC analysis indicates sea ice. The greatest areal extent of the conflicts occur in the southern ocean although the difference in SST between the original and purified SST analysis is not large. In the Arctic, we find more unfrozen points in the OISST where the SSTs are considerably warmer that sea ice.

To reduce such conflicts the original daily OISST were adjusted in the internal sea ice margins to be reflective of an ocean-only analysis. This "purification" process works as follows:

• The first and last five rows on the global grid (85-90 deg) were set to the sea ice temperature of -1.8 C.
• The remaining sea ice points were set to undefined
• The undefined points were filled in using a Cressman scan analysis procedure dubbed the "weaver" at NCEP. This is the same routine used to fill or define SST values over land and insures a reasonable transition across the land/sea boundary.

Finally, there are three missing days in the daily-interpolated SST in 1995: 6,12, and 17 November.

Masks

The land sea mask is the same as used in the NCEP OISST. This mask has more ocean points than would be typically used in a model, but has the advantage of allowing the SST analysis to have more influence at the land/sea boundary -- a desirable feature when interpolating to coarser resolution models.

The glacier mask that came from the NCEP reanalysis is used primarily to differentiate open ocean from the permanent ice shelves of Antarctica. The mask was originally 2 deg, but was adjusted manually to be more representative of higher resolution ice shelves as seen in [Gloersen, 1992 #41].

A graphic of the mask may be found by clicking here.

Creation of the SIC BCSOBS

Although the three SIC source data sets have different sampling frequencies and horizontal resolution, these data sets are all based on passive microwave observations from the SSMR and SSM/I instruments aboard the Nimbus 7 and DMSP satellites. Further, they use a common algorithm to convert the radiation to sea ice concentration.

The two basic data sets are identified by their creators A. Nomura (JMA and ECMWF, see [Nomura, 1995 #32]) and R. Grumbine (NCEP, see [Grumbine, 1996 #34]) and the final data set for AMIP II will be called "AMIP II." For the Grumbine sea ice, the documentation may be found in (http://polar.ncep.noaa.gov/seaice/icegroup.html). Also, click here for an excellent starting point on sea ice.

The original satellite data was on a ~25 km polar stereographic grid and was then interpolated to a lat/lon grid for use in models. Nomura applied extensive quality control of the lat/lon sea ice concentration to eliminate bad and missing data. He choose a weekly time interval to insure sufficient observations in each grid box as the SSMR sampling was less than the SSM/I. Whereas Nomura started with SIC, Grumbine uses the radiation data directly and applies quality control to both the source data and SIC product. More significantly, Grumbine performs his analysis daily.

Intercomparison during the overlap period

We first compare the two data sets in the one month where they overlap - December of 1991. Grumbine uses an updated algorithm for the radiance to sea ice conversion and more extensive quality control. Close correspondence would indicate the resulting sea ice analysis is dominated by the observations and not the process (i.e., the "what" versus the "how"). In the Northern Hemisphere we find very good agreement (I have only averaged and rectified the grids at this point), even when looking in the difference field. However, the Grumbine field has a slightly higher net concentration, particularly in Hudsons Bay, but these differences are well within the error bounds (3-10%).

The agreement in the Southern Hemisphere is less good and we find consistently lower concentrations in the difference field. Also, note the differences in at the glacier/land - ocean interface - a consequence of different masks. Nevertheless, the difference are still within error bounds and, as will be shown later, no large discontinuites were discovered. Thus, we conclude that the Grumbine and Nomura data are likely comparable and consistent even though the Grumbine data is expected a priori to be of higher quality because of improved observations and processing.

Data Adjustments

While the intercomparison showed reasonable agreement between the data sets, some differences in the final product used by reanalysis remained due to how the polar stereo data was interpolated to the 1 (0.5) deg lat/lon grid and different land masks. Thus, some minor adjustments were required to generate the final AMIP II SIC monthly means.

The process is outlined below:

1. Generation of the monthly mean SIC

• average the daily SIC data to form the monthly mean. For Nomura the "daily" data comes from an interpolation of original weekly analyses to daily values as in the NCEP reanalysis. In contrast, the Grumbine data is a true daily analysis.
• set values >= 0.98 to 1.0 to rectify differences in how high ice concentration is treated (Nomura 0.99 and Grumbine 1.00).
• If undefined, set the unobserved values near the North Pole (85-90) to 1.0.
• move the data to the standard 360x180 1deg grid where (1,1) = 0.5E, 89.5S by:
• area-weighted regridding of the 0.5 deg data
• shifting in longitude (Nomura (1,1) = (179.5W, 89.5S) ; Grumbine (1,1) = 0.5W, 89.5N)
• fill the undefined points in the Nomura data using successive regridding. This procedure extends defined data into the undefined regions and gives smooth transitions at the coastlines. The purpose is to provide reasonable values when interpolating to a coarser (generally) model grid. The regrid extrapolation does not change any defined data.
• generate masks where SIC > 0.10, 0.15, 0.35, 0.55, 0.70 (1=sea ice, 0 = open ocean, undefined values at land or glacier points).
• generate SIC from the GISST2.2 data using the same algorithm as above for intercomparison purposes.

1. Calculate "sea ice extent"

• Sea ice extent is defined by [Gloersen, 1992 #41] as the surface of the earth with SIC > 15 % and expressed as millions of km² . We calculate sea ice extent in the same (roughly) 12 areas as [Gloersen, 1992 #41] for intercomparison purposes and to verify the data is reasonable. A similar analysis would have uncovered the more egregious errors in the AMIP I SIC in the Southern Hemisphere.

Properties of the SIC BCSOBS

An interesting technical feature of the data were infrequent, large areas of very low SIC (2% and lower) in both sets, e.g., December, 1990 (Nomura) and June, 1995 (Grumbine). For models that use SIC instead of a mask and that might respond to very low values, a check would be needed. I did not filter these points out because they could be realistic in the ice margins and in the regions of low ice concentration near the Ross Ice Shelf as seen in the 17-year mean for February (summer time). Further, such a filter algorithm is not grossly obvious and might not be necessarily appropriate for all models.

To detect trends and temporal consistency we examine the sea ice extent calculated in a similar way as in the Gloersen et al. 1992 NASA SSMR sea ice atlas. Because of different land/sea masks and crude area definition there are differences between the AMIP II sea ice extent and that in the NASA atlas. However, comparing the Arctic time series to that on p. 114 of the atlas, we see excellent agreement in the low ice months but about 15 % greater extent in the winter. This difference is mostly a consequence of the mask (I have few land points). Because we need data at all points for interpolation to the model grids, this discrepancy is considered acceptable. The key point is that character of the time series is similar and no untoward trends are found in the AMIP II data. In contrast, the GISST2.2a data shows a serious negative trend.

Another notable feature is the large increase in December of 1995 in both the Arctic and Antarctic time series. This is the point where Grumbine shifts to a 0.5 degree analysis and further investigation and more data are needed to understand if these data are biased.

The table below provides time series in the other Gloersen et al. areas and gives some comments on the plots.

Intercomparison of GISST2.2a and AMIP II sea ice extent

 
 

Sea / Region	Comments
Bering Sea (Arctic)	Positive trend in AMIP II ice, GISST no trend, but more ice in 1994-96 period in AMIP II.
Labador Sea (Arctic)	Substantial, ENSO-like interannual variation in both data sets, positive trend in AMIP II ice
Greenland Sea (Arctic)	Negative trend in GISST ice, increase in the 94-96 period, more ice in GISST
Berents Sea (Arctic)	Negative trend in GISST, slight negative trend in AMIP II, interesting interannual variation
Sea of O / Japan (Arctic)	Slight negative trend in both data sets, more ice in AMIP II
ARCTIC OCEAN	Slightly more ice in AMIP II, but strong negative trend in GISST
Indian Ocean Sector (Antarctic)	AMIP II more consistent, big jump in ice in the GISST data around 1989, large change in summer time ice in 1996 (shift to 0.5 degree analysis)
Western Pacific Sector (Antarctic)	Large interannual variation, less trend in AMIP II
Ross Sea (Antarctic)	Large positive trends in both data sets, similar 1996 summer ice feature as in the IO sector
Amundson Sea (Antarctic)	Strong interannual variation in both sets, slight but opposite trends
Weddell Sea (Antarctic)	Positive trend in GISST because of greater ice in 1990's
ANTARCTIC OCEAN	More ice and a big jump starting in 1989 in the GISST data leading to a positive bias in GISST

Despite some of the "interesting" features (occasional large areas of very low ice concentration and the differences in the 0.5 versus 1.0 deg analyses), the basic finding is that the AMIP II SIC, because of good temporal consistent and better data sources, is more suitable for the 17-year integration than the GISST2.2a SIC.

One final note is that the Grumbine SIC, after the change to daily analyses (December, 1995), has values over Antarctica whereas the Nomura SIC and the early weekly analyses do not. Thus, plots of SIC after this change will show quasi realistic values, but the regrid interpolation does not extend values (100%) to the South Pole and they are remain zero as intialized. This "feature" should have no affect on models with reasonable land/glacier masks.

Creation of the SST BCSOBS

Unlike SIC, where the concentrations were derived from the same type of observing platform, two different sources for data were used to construct the AMIP II SST BCSOBS. These two sources differ in:

• availability of satellite data
• treatment of SST in the ice margins
• EOF reconstruction in the tropical-midlatitude oceans

and the properties of the resulting SST time series did, not unexpectedly, show large changes as the source data changed. The processing will be described in the sections below according to the major decisions in adjusting the data for AMIP II.

Using the original monthly means for the GISST2.2a data instead of the daily-interpolated fields from reanalysis (GISST2.2)

The original GISST2.2 data are monthly means and these means were linearly interpolated to create daily values for the NCEP reanalysis during the pre-OISST period (November, 1981-present) by assuming that the monthly means were valid at the midpoint of the month. Full consistency with the NCEP reanalysis would require averaging these daily interpolated fields, but there were two problems with this approach: 1) an error in the ice margins discovered at the UKMO; and 2) damping of the annual cycle (AC). Thus, we used the corrected (GISST2.2a) monthly means.

To illustrate the AC damping problem, examine the difference between monthly means averaged from daily values and the original monthly means by clicking here. Note the positive differences on the order of 0.5 C but sometimes > 1 C in the summer hemisphere midlatitudes. This implies a substantial (> 10%) bias in the amplitude of the annual cycle in the midlatitudes. A similar but smaller problem was found in the Southern Hemisphere. In this case, note the strong cool anomalies in the Gulf Stream and Kuroshio currents.

Rectify the SST to SIC

As noted earlier, different SIC data sets were used in the source SST data and, as discussed, the first adjustment was to "purify" or preprocess the source SST data (daily or monthly mean) in its own sea ice zones. However, because of different values of SST in the SIC mask regions (SIC > 55 %), averaging of the daily OISST means and different land/sea masks; different SSTs were found in SIC margins. Hence, the monthly means were adjusted to the same SST in the SIC mask regions, and in the SIC margins (10 % < SIC < 55 %) to insure smooth, and consistent transitions from an ice free to a frozen surface.

We first define two SSTs :1) SST_{SIC Mask} = 271.38 K ; and 2) SST_{SIC Mask +} = 271.48 K for the rectification. In the final SST, all points in the SIC mask region will equal SST_{SIC Mask} , but first all SST below SST_{SIC Mask +} are set to SST_{SIC Mask +} to eliminate excessively cold points. Next, all points in the glacier regions and the SIC mask region are set to SST_{SIC Mask} . The remaining land-only points (not sea and not glacier), and the points in the sea ice margin are set to the undefined value. Thus, the SST field at this point equals: 1) SST_{SIC Mask} in the sea ice mask region and at glacier points; 2) the original value where SIC > 10%; and 3) undefined over land and in the sea ice margin. We next apply the "weaver" algorithm as described in the SST preprocessing section to set these undefined points to give smooth transition in the margins. However, the SST in the ice free regions (SIC > 10%) may still be excessively cold if the point was classified as sea ice in the original data, but in the adjusted SST it will at least not be frozen (SST_{SIC Mask+}).

Differences in the annual cycle in the ice margins

The GISST2.2a and OISST data treat the water temperature in the sea ice zones very differently. GISST2.2a sets the SST in the SIC > 0 region using an empirical relationship based on observations. In contrast, the OISST sets the SST to a constant on the sea ice mask (SIC > 50 % not 55 % as in ERA) zone and the analysis procedure defines the SST in the marginal ice zones (0 < SIC < 50 %). Thus, the SST could conflict not only from differences in SIC, but in how the SST was treated in the ice margins and a serious difference was uncovered. Examining the zonally averaged SST with the annual cycle removed shows anomalous warm water in the marginal sea ice regions around 80 N and 70S only during the GISST2.2a time.

The annual cycle (AC) was then calculated for the three-year period before (GISST2.2a AC) and after (OISST AC) the change to OISST. The difference field is striking, e.g., see the January field. While there is a real change in the AC in the tropics and midlatitude oceans, a persistent warm bias was found in the high latitude ice margin. Use the table below to view a particular month:

Intercomparison of the GISST2.2a AC 7812-8111 and the OISST 8212 - 8411

 
 

JAN

FEB

MAR

APR

MAY

JUN

JUL

AUG

SEP

OCT

NOV

DEC

To eliminate the large differences in the AC in the data sets, the GISST2.2a AC was adjusted to be consistent with the OISST. We are not suggested that the OISST AC is necessarily better (i.e., more accurate). Rather, temporal consistency with the dominant source of the SST (OISST) necessitates the adjustment. The modification is to simply replace the 3-year GISST2.2a AC with the 3-year OISST AC (i.e., the AC based on the three years before and after the change to OISST). However, the replacement is weighted so that only the high latitude AC is changed. Specifically, no adjustment equatorward of 55 deg and full replacement poleward of 65 deg. An example of the correction at a specific longitude and time graphically shows the large differences. Note that some points are below SST_{SIC Mask.} . This illustrates how the adjustment distorted the SIC-SST consistency. Thus, we performed the SIC-SST rectification again to bring the SIC and SST back in synch, but only for the GISST2.2a period. Also, the overcorrection only occurred at only a few points.

The corrected version of the original zonal average SST shows that the AC replacement has yielded a consistent time variation in the high latitudes without affecting the tropics and midlatitudes. It is also interesting to note the apparent warming trend near ice margins from 1991-1995 in the Northern Hemisphere. This trend is not clearly linked with the corresponding SIC trends in the AMIP II SIC. The big increase in sea ice extent in 1996 and the cooler anomalies is interesting, but may be questioned because of the change in the SIC analysis grid.

The small SIC-SST intercomparison above suggests that the SST should be reanalyzed using the more consistent AMIP II-type SIC and that such a reanalysis should explicitly link the SIC and SST in a more physical/observational manner (the explicit SIC-SST relationship in the GISST2.2a data is a step in this direction). Despite the lack of a direct connection between the AMIP II SIC and SST observations, we believe these data meet the requirements of the experiment. That is,

• Consistency with the NCEP and ECMWF reanalyses
• Consistency in time
• A physical consistency between SST and SIC

While SIC-SST reanalysis (i.e., a comprehensive near surface oceanographic reanalysis) is clearly needed, these AMIP II BCS observations are probably the best available for climate model integrations during the 17-year AMIP II period.

AMIP II BCSOBS Graphics

A full set of graphics (3 areas (global, nhem and shem) * 2 fields (SIC and SST) * 2 types (mean and anomaly) * (208 months + 12 months of climatology) ~ 2600 plots) is available by clicking here (or on the section title). This page is a form that lets you "point and click" to select a plot.

AMIP II BCSOBS Data

There are several ways to access the data. The recommended path is by having PCMDI generate the BCS and the BCSOBS on your grid. See AMIP Sea Surface Temperature and Sea Ice Concentration Boundary Conditions where you will find links to our generation process as well as links to the data on the 1° grid (http://www-pcmdi.llnl.gov/projects/amip/amip2/AMIP2EXPDSN/BCS/amipobs_dwnld.html) in a variety of formats. This document also gives a complete description of the AMIP II BCS.

Conclusion

We hope the interactive nature of this document and the extensive plots will be useful in developing a clearer understanding of the data. No analysis and discussion of the data can substitute for a modeler / user's direct examination.

Your findings and discussion are needed to make this document a true community resource. Please forward them to Mike at fiorino@llnl.gov.

---

Last update 12 September 1997

LLNL Disclaimers

UCRL-MI-125597