the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Data assimilation schemes for ocean forecasting: state of the art
Abstract. Data assimilation is a process for integrating models and observations into comprehensive and reliable estimates of the ocean state. It is used to produce near-real time initial conditions (analyses) from which ocean forecasts are produced and to generate reconstructions of the past state of the ocean (reanalyses). Here we provide an overview of the methods currently used in ocean systems for assimilating satellite and in-situ observations, together with a brief review of methods being developed which will be implemented in future operational systems, including the use of machine-learning techniques that provide a way to improve their efficiency. A list of data assimilation software used by most of the global and regional operational ocean forecasting systems is provided, together with its availability. A discussion of practical considerations for employing data assimilation software and techniques operationally is also given, including the types of observations which are commonly used, and the implementation choices made by existing operational systems at global and regional scales is summarized.
- Preprint
(858 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (extended)
-
CC1: 'Comment on sp-2024-20', P. Sakov, 24 Sep 2024
reply
This is a quick comment regarding the top figure on p. 11 (l. 275).
On 29 June 2022 Australian Bureau of Meteorology has transitioned its operational global ocean forecasting system from OceanMAPS v3.4 to OceanMAPS v4.0. OceanMAPS v4.0 is a hybrid EnKF/EnOI system with 48 dynamic members and 144 static members. OceanMAPS v4.0 is performing beautifully since then and is about to be upgraded to v4.1 with a 1-day cycle. Therefore, the entry for Australia on the above figure needs to be changed to "BoM, Australia. EnKF-C | EnKF. 1/10 degree | 3 days. SST, SLA, T/S."
Citation: https://doi.org/10.5194/sp-2024-20-CC1 -
AC1: 'Reply on CC1', Matthew Martin, 24 Sep 2024
reply
Thanks for the correction on this Pavel. I've updated the figure to replace EnOI with EnKF and adjusted the text starting line 242 to read:
"In general, the global systems use somewhat simpler DA algorithms (though they are still complex in their implementation of those algorithms) than the regional and coastal systems, the exception being the BoM system which uses a hybrid-EnKF with 48 dynamic members and 144 stationary low-mode members (Brassington et al., 2023). Many global forecasting groups use a 3DVAR-FGAT algorithm (Barbosa Aguiar et al., 2024; Zuo et al., 2019; Cummings and Smedstad, 2013; Storto et al., 2016; Ravichandran et al., 2013) with some groups using a SEEK filter or LESTKF with a static ensemble (Lellouche et al., 2018; Smith et al., 2016; Li et al., 2021)."
And I've included the reference:
Brassington, G. B., Sakov, P., Divakaran, P., Aijaz, S., Sweeney-Van Kinderen, J., Huang, X., and Allen, S.: OceanMAPS v4. 0i: a global eddy resolving EnKF ocean forecasting system, in: OCEANS 2023-Limerick, IEEE, 1–8, https://doi.org/10.1109/OCEANSLimerick52467.2023.10244383, 2023.
Citation: https://doi.org/10.5194/sp-2024-20-AC1
-
AC1: 'Reply on CC1', Matthew Martin, 24 Sep 2024
reply
-
RC1: 'Comment on sp-2024-20', G. C. Smith, 25 Oct 2024
reply
Review of “Data assimilation schemes for ocean forecasting: state of the art” by M.J. Martin et al.
General comments:
The international landscape of operational oceanographic forecasting systems (and the data assimilation schemes used) has evolved considerably over the last ten years. It is thus timely that Martin et al. provide this overview and description of the different methods in place and how they are implemented globally and regionally. The review provides a thoughtful and accurate description of the various methods applied, together with comments regarding particular challenges and benefits of each. Overall, I find the paper provides a useful reference regarding the current state of ocean data assimilation systems and their implementations. Below I’ve provided a few comments the authors may want to consider to further improve the breadth and impact of the paper.
Main points
- It is not clear to me reading this paper to what readership it is intended. The introduction provides a quite high-level background on operational oceanography. This is then followed by a description of methods, which is often quite technical (e.g. reference to model adjoints, etc..). It ends without any kind of summary, discussion or perspective toward the future. It reads as though it is a chapter in book, but I didn’t see any reference to a Special Issue. I’m not familiar with this journal and it appears to organize papers into chapters somehow. Regardless, it would be helpful that the paper is either more self-contained or references the contextual pieces.
- The introduction is quite general while the rest of the paper digs into some fairly detailed comments about the methods. There is little, however, to provide context regarding the particular DA challenges for ocean prediction and its history. Rather the paper focuses almost exclusively on the methods themselves.
- For example, ocean DA methods have evolved in large part from those developed for Numerical Weather Prediction. Some mention to this effect would be appropriate.
- The impact of particular challenges of ocean DA on the methods used would be relevant. For example, the spatial and temporal scales of the ocean are quite different than for the atmosphere. The observations available are also quite different with satellite observations only providing surface information. This affects both the cost / benefit of using ensembles as compared to higher model resolution as well as limitations on constrained scales due to information content of satellite altimetry. While I understand the focus of the paper is on DA methods, the choice of method depends on the model used and what observations are available. Some discussion on this point could be made in Section 5 on Ocean Observations as well.
- Another area I felt was underrepresented was a discussion of errors. DA aims to provide a best estimate of the system state based on estimates of model and observational errors. In this sense, it could be helpful to outline what the key errors are that are being accounted for (e.g. due to model physics, resolution, forcing, intrinsic variability). Some comment regarding the treatment of bias and representivity error would also be relevant.
- I found the lack of any kind of summary, conclusions or discussion section to be quite unusual. Perhaps some historical context to note how quickly the landscape of operational oceanographic systems has been changing and how developments in satellite observing systems (e.g. SWOT, surface current retrievals) and ML methods may lead to further rapid developments. Challenges associated with constraints on the submesoscale circulation (which affects many applications) could also be mentioned as a limitation of current ocean DA.
Minor comments
Intro line 37 “important processes can be constrained.” It is not really a process that is constrained but rather the associated variability in essential ocean variables.
Line 59: “When the ocean model is considered perfect…”, I would specify “When the ocean model AND ATMOSPHERIC FORCING is considered perfect…”
Line 70: “…adjoint model, governed by the adjoint equations to the ocean tangent linear model”. These terms are introduced without any explanation. Some background or at least a reference would help readers to follow.
Citation: https://doi.org/10.5194/sp-2024-20-RC1 -
CC2: 'Comment on sp-2024-20', Lars Nerger, 11 Nov 2024
reply
Dear Authors,
looking through the manusript I found a few aspects I like to comment about:
1. line 91 mentions 'Reduced-order extended Kalman filter (ROEK)'. Actually, this very early development never seemed to attain any relevance. As such it's surprising to find this in a manuscript intending to become a reference paper. Actually, the concept of error-subspace DA (linked to SEEK) seems to be more relevant (we also find this e.g. in the publications about 'unstable subspace'). Considering the error-subspace SEEK and EnKFs are also rather similar as both represent an error-subspace (but the linearized dynamics used in the SEEK filter tend to yield worse performance as was by demonstrated by Nerger et al. (2005). A Comparison of Error Subspace Kalman Filters, Tellus A, 57A(5), 715-735, doi:10.1111/j.1600-0870.2005.00141.x)
2. In line 92 the manuscript cites 'Hoteit et al., 2018' for the EnKF. This is a book chapter that was not peer reviewed. Obviously there are relevant original peer-reviewed articles (e.g. Evensen, 1994, Burgers et al, 1998, Houtekamer & Mitchell, 1998) about the EnKF, but also peer reviewed articles that provide an overview of the state of the art of EnKFs (e.g. Vetra-Carvalho et al., Tellus A, 2018, doi:10.1080/16000870.2018.1445364). Scientific diligence would perhaps rather call for citing such peer-reviewed references.
3. Regarding particle filters (~line 108), van Leeuwen (2015) is cited. Actually, since then there was progress in the developments. Van Leeuwen et al. (2019) Particle filters for high-dimensional geoscience applications: a review. Quarterly Journal of the Royal Meteorological Society, 145, 2335-2365, doi:10.1002/qj.3551 gives a more comprehensive review including examples from high-dimensional applications of the PF (in the atmosphere, not the ocean).
4. The description of PDAF in Table 1 is not fully correct: PDAF uses features of Fortran 2003. Further, PDAF targets not only ensemble DA, but also includes 3D-Var. For the link to the website, I recommend to only state 'https://pdaf.awi.de', the sub-directory '/trac/wiki' is an automatic forwarding (and while the domain will certainly be conserved the sub-directory might disappear in case that we update the software running on the server). It would be good if these aspects could be taken into account.
5. Figure 2 on operational regional and coastal ocean DA systems seems to be rather incomplete and it's not clear why the particular systems were chosen. E.g. in Germany a regional system for the North Sea and Baltic Sea is run (e.g. described in Bruening et al., Hydrographische Nachrichten 118 (2021) 6-15, doi:10.23784/HN118-01) This system uses a fully dynamic LESTKF with PDAF. Also next, to the SMHI-System that is mentioned, there is the operational system for the Baltic Sea of the EU Copernicus program run by the Baltic Monitoring and Forecasting Center (BAL-MFC, see e.g. https://marine.copernicus.eu/; there doesn't seem to be a proper peer-reviewed article, but the center descriptions and the documentation of the data products), but also the Danish Meteorological Institute seems to run their own system. Apart form this NMEFC also runs an ensmeble-based system for sea ice in the Arctic ocean (e.g. Liang, X., Z. Tian, F. Zhao, M. Li, N. Liu, C. Li (2024) Evaluation of the ArcIOPS sea ice forecasts during 2021–2023. Front. Earth Sci. 12, 1477626 doi:10.3389/feart.2024.1477626). Actually, all these systems use PDAF leveraging commonalities
Best regards,
Lars NergerCitation: https://doi.org/10.5194/sp-2024-20-CC2
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
204 | 63 | 141 | 408 | 3 | 3 |
- HTML: 204
- PDF: 63
- XML: 141
- Total: 408
- BibTeX: 3
- EndNote: 3
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1