SODA logo

Background



History The origins of the Simple Ocean Data Assimilation (SODA) renalysis go back to the early 1980s as an outcome of an effort to merge a wide variety of observations from oceanographic experiments when I was a postdoc with Alan Robinson at Harvard (e.g. Robinson et al., 1984, Robinson et al., 1986). In the mid-1980s after I moved to University of Maryland I became involved in the joint US/French FOCAL/SEQUAL program in the tropical Atlantic. FOCAL/SEQUAL had also collected a wide variety of observations of different types. The first implementation of the SODA data assimilation package began as an effort to merge together F/S data and relate them to atmospheric variability. This work was done in collaboration with Eric Hackert. Several papers were published in 1989 and 1990 describing this work (Carton and Hackert, 1989; Carton and Hackert, 1990; Raghunath and Carton, 1990). Among the issues addressed in these papers is the challenge of combining velocity observations with temperature and salinity observations under the constraints of geostrophy.

The domain of our interest, and thus SODA, gradually expanded to a global domain during the 1990s and we carried out a number of studies with the (then new) satellite SST, and SSH data sets. The domain expansion was made practical by the release of the NCAR/NCEP global atmospheric reanalysis (Kalnay et al., 1996). At the same time Gennady Chepurin and I were developing SODA (I named it in analogy to the Simple Biosphere Model being developed by Piers Sellers at the same time) we were also talking to the developers of related ocean efforts at NCEP (Ants Leetmaa, David Behringer, and Ming Ji), and GFDL (Tony Rosati) as well as meteorologists at Goddard and NCEP. Thus, some of the ideas implemented in SODA have several parents. For example, I borrowed Bloom's Incremental Analysis Update and the idea of flow-dependent covariances from the Goddard GMAO group.

We carried out the first intercomparison of reanalysis reanalysis products in the late 1990s, showing that all were doing alright in the tropical and North Pacific and generally better than the no-model analyses (Chepurin and Carton, 1999). This early work was funded by a joint NSF grant with Syd Levitus, who helped us sort out the global hydrographic data set, followed by a grant with Eugenia Kalnay. The work on the quasi-global version of SODA culminated in two papers in 2000, one describing the algorithm and the other presenting a number of data comparisons (Carton et al., 2000a,b). A number of versions of this reanalysis were released, perhaps the latest of which was SODA version beta23. The original work on bias correction was also done with this model (Chepurin et al., 2005).

The original ocean model was based on GFDL MOM2 physics and had a rather coarse one degree (initially even coarser) numerical grid. A second problem that became apparent is that the surface forcing used to drive the model had some spurious features including longterm trends introduced by changes in the atmospheric observing system. A collaboration with Julie McClean of Scripps provided a copy of a global Parallel Ocean Program POP v1.4 model with 40 levels in the vertical and a 0.4x0.25 degree displaced pole grid (25 km resolution in the western North Atlantic. The problems with the surface forcing were addressed by a switch to ECMWF-ERA40 forcing. Testing of the new model proved to be quite laborious and consumed considerable effort by Xianhe Cao and Ben Giese who had joined us in the early 1990s as a postdoc and left for a faculty position at TAMU in the mid-1990s. Ben and Cao both wrote interfaces between the model and the assimilation codes and we eventually went with Ben's version. Ben also made the choices regarding the interpolation of the output onto a uniform 1/2-degree grid. Finally, he arranged for distribution of the data sets by other websites. The data assimilation coding, including parallelization, and the basic data sets were the provence of Gennady Chepurin.

The first version of the reanalysis based on the POP model to be released (briefly referred to as SODA POP) was SODA1.2 which still used NCEP reanalysis winds. This version is also described in the methodology section below. Evaluation of the results were a joint effort of the UMD group (Cao, Chepurin, and Carton). The second, SODA1.4.2, used the ECMWF ERA40 winds. This reanalysis was described in Carton and Giese (2008), which summarized many of the developments up to that point.

The next NSF grant supporting SODA was a collaborative grant with Ben Giese with the focus of developing more sophisticated data assimilation and exploiting the atmospheric reanalysis work of Gil Compo and colleagues at NOAA ESRL to extending the time period of the data assimilation analysis back to the beginning of the 20th century. Several versions of this analysis were carried out by Ben including a pure simulation (SODA2.2.0 and an analysis using the same data assimilation (SODA2.2.4). This experiment was described in a separate paper (Giese and Ray, 2012).

Some additional experiments were made to test the impact of bias in the hydrographic profiles and the problem of historical data sampling (Giese, et al., 2011; Carton et al., 2012), however the assimilation code remained the same. Most recently Ben has been carrying out additional experiments with Compo and colleagues in which the assimilation is mostly stripped out except SST. Multiple analyses are forced by separate members of the ensemble of atmospheric analyses (SODA-si).

The SODA data assimilation development at UMD, meanwhile, has been moving in parallel with developments in atmospheric assimilation. Steve Penny, who was an applied math student at UMD (jointly advised by Carton and Kalnay) completed his dissertation on an implementation of an ensemble Kalman Filter in SODA (Penny et al., 2013). While results were encouraging, the cost of a full ensemble filter is constraining. As a postdoc working jointly between UMD and NCEP Steve has extended his work to develop a hybrid filter which retains many of the appealing features of the ensemble filter, while remaining practical. Results from an Observing System Simulation Experiment in comparison with the NCEP GODAS system is in review now (Penny et al., 2015). Experiments with historical data are being carried out as of this writing.

One major problem areas has been the quality of the analysis at high latitudes. The most recent NSF grant in collaboration with Mike Steele at APL and Sirpa Hakkinen at Goddard has particular emphasis on this region and issues related to tracking sea ice and the freshwater budget. The shift to a new ocean model (MOM5 physics) that includes active sea ice is being done to address some of these issues and is being carried out in collaboration with GFDL, NCEP, and Goddard/GMAO.

Methodology The Optimal Interpolation code, developed for the tropical Atlantic, was the basis of the code extended to the global ocean (Carton et al., 2000a,b). SODA version Beta23 was perhaps the earliest version released to the community. The non-eddy permitting model that was used in this reanalysis suffered from serious biases. To address this problem we began a collaboration with Dick Dee to develop a bias-aware filter, eventually described in Chepurin et al. (2005). From SODA version Beta23 through SODA2.2.4 the assimilation code remained largely fixed with the major improvements coming from use of an improved ocean model (the shift to an eddy-permitting version of the Parallel Ocean Program code), better and more extensive ocean data sets, and improved surface meteorology. The first analyses using this system were: SODA1.0 and SODA1.2 using the NCEP/NCAR reanalysis, and SODA1.4.2 using ECMWF ERA40 surface meteorology.

Eventually we reached the point where the assimilation scheme needed updating. An applied math graduate student Steve Penny (co-advised by Eugenia Kalnay) took up the challenge. Steve implemented the Local Ensemble Transform Kalman Filter in SODA and presented a number of comparisons in his dissertation (Penny et al., 2014). Steve has continued this work with us in collaboration with David Behringer of NOAA/NCEP, developing a hybrid gain filter (Penny, 2014) which retains advantages of both schemes. Steve has carried out a series of observing system simulation experiments (Penny et al., 2015) and more recently with the historical observations the results of which look promising.

References: mostly in: www.aosc.umd.edu/~carton/carton/ref.html

Table 1: list SODA Reanalysis Experiments based on the POP ocean model.

Name

Winds

Time

Model

Obs

Notes

1.21

NCEP

1948-2003

POP1.4

WOD01

 

1.4.0 (2)

ERA40

1958-2001

POP1.4

simulation

These widely circulated.  The  most noticeable problem we’ve found is the occasional appearance of spurious cold anomalies at thermocline depths.

1.4.2 (2)

ERA40

1958-2001

POP1.4

WOD01

1.4.3 (2)

QSCAT

2000-2005

POP1.4

WOD01

1.4.4

ERA40

1992-2001

POP1.4

WOD01/ALTIM

2.0.2 (3)

ERA40

1958-2001

POP2.0

WOD05

 

2.0.3 (3)

QSCAT

2002-2005

POP2.0

WOD05

Contains some bad ARGO data in the Atlantic

2.0.4 (3)

QSCAT

2002-2007

POP2.0

WOD05

2007 is bad

2.1.6 (3)

ERA40/QSCAT

1958-2008

POP2.1

WOD09

2007 fixed, also T/S variations should be more accurate

2.1.0

ERA40/QSCAT

1958-2007

POP2.1

WOD05

2.1.0, 2.1.2-4 all are missing mixed layer SST updating (part of the experiment design).  These experiments are designed to explore the impact of fall rate corrections

2.1.2

ERA40

1958-2001

POP2.1

WOD05

2.1.3

QSCAT

2002-2007

POP2.1

WOD05

2.1.4

ERA40/QSCAT

1958-2004

POP2.1

WOD05

2.2.0

C20R-2

1890-2007

POP2.1

simulation

These experiments are part of an ongoing examination of 20th century climate.

2.2.4

C20R-2

1890-2010

POP2.1

WOD09

1Described in Carton et al. (2005).  Available at monthly resolution  from: iridl.ldeo.columbia.edu. The potential user is reminded that the analysis is carried out on a non-Mercator coordinate grid and then mapped onto a regular Mercator grid (0.5 deg resolution). This mapping, done for the convenience of the user, means that you need to be careful if you compute certain kinds of integrals such as meridional heat transport.

2Described in Carton and Giese (2008).  1.4.2 and 1.4.3 are available at monthly resolution through apdrc.soest.hawaii.edu and also iridl.ldeo.columbia.edu

3  Many products up to and including 2.2.4 are available at monthly resolution through apdrc.soest.hawaii.edu and several also through https://iridl.ldeo.columbia.edu.  At each site you will need to search for SODA (I can't point you to the direct link since this changes).. Finally, a few products, including the monthly regridded version of 2.2.4 are available from UMD through https://dsrs.atmos.umd.edu/DATA

4With fall rate correction as proposed by Levitus et al. (2009)

5With fall rate correction as proposed by Wijffels et al. (2008)