the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
LFS-SEnDA: a global ocean forecast system with an efficient Sliding-Ensemble-based Data Assimilation scheme
Abstract. This study develops an efficient Sliding-Ensemble-based Data Assimilation (SEnDA) scheme and integrates it into the LICOM Forecast System to establish LFS-SEnDA. In SEnDA, time-dependent sliding ensemble samples are used to represent temporally varying background-error covariance structures. Meanwhile, SEnDA has incorporated a parallel-friendly analysis procedure and an accelerated solver to improve computational efficiency. Compared with the conventional Ensemble Optimal Interpolation scheme, SEnDA improves analysis quality, particularly in eddy-rich regions, and achieves about a 30-fold speedup in analysis step. Within the IV-TT Class-4 assessment, LFS-SEnDA ranks among the leading candidate global ocean forecast systems and achieves the highest composite score based on sea surface temperature, sea level anomaly, upper-ocean temperature, and salinity over forecast days 1–6. At a 1-day lead time, LFS-SEnDA achieves global RMSEs of approximately 0.38 °C for sea surface temperature and 0.06 m for sea level anomaly. These results indicate that LFS-SEnDA provides an efficient and scalable framework for improving multi-variate ocean state estimation and global ocean forecasting.
- Preprint
(2453 KB) - Metadata XML
-
Supplement
(1285 KB) - BibTeX
- EndNote
Status: open (until 25 Dec 2026)
- CC1: 'Comment on egusphere-2026-4602', Hao Luo, 12 Nov 2026 reply
-
CC2: 'Comment on egusphere-2026-4602', Chuanyu Liu, 20 Nov 2026
reply
This study develops a Sliding-Ensemble-based Data Assimilation (SEnDA) scheme where time-dependent sliding ensemble samples are used to represent temporally varying background-error covariance structures. The SEnDA is incorporate to the high-resolution (1/10°) LICOM Forecast System. Compared with the conventional Ensemble Optimal Interpolation scheme, the SEnDA improves analysis quality, particularly in eddy-rich regions. Within the IV-TT Class-4 assessment, LFS-SEnDA ranks among the leading candidate global ocean forecast systems. This study represents new efforts in improving the high-resolution global ocean forecast. The techniques, model setups, experimental comparisons are properly described. The results are well illustrated. I am glad to see such new improvements, and I would like to suggest acceptance of this article, after the following minor concerns are addressed. These minor comments might be helpful for the authors to clarify the manuscript.
L14, suggest adding “(LFS)” after “LICOM Forecast System”
L48-50: provide more details to explain why the EnOI and SEEK methods weakens the flow dependence, so that the readers can see the advantages of the SEnDA.
L97: give more details on the reconstructed profile and reprocessed gridded T/S products.
L107: when is f(t) prescribed 1 and when it is set 2?
L115: how to select samples for the current assimilation date from historical simulations?
Section 2.1.2, better to give more weight on the description of SEnDA, particularly, the “sliding’ feature needs to be emphasized and clearly addressed.
L198: what does “each consisting of the first day of each month” mean? Why only the first day?
L200: should the sliding strategy have employed 300 (=30*10) samples? Why 120?
Eq. 7 and L231, what is P?
Figure 5: clarify what the results in b and d are? From the cycling assimilation-forecast system? What is the x axis? Cite Fig.6b and d individually in the text to make correspondence between text and figure.
Citation: https://doi.org/10.5194/egusphere-2026-4602-CC2
Data sets
LFS-SEnDA-Dataset Yanzhi Zhou et al. https://doi.org/10.5281/zenodo.23165690
Model code and software
LFS-SEnDA-Code Yanzhi Zhou et al. https://doi.org/10.5281/zenodo.21738114
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 276 | 93 | 30 | 399 | 53 | 37 | 37 |
- HTML: 276
- PDF: 93
- XML: 30
- Total: 399
- Supplement: 53
- BibTeX: 37
- EndNote: 37
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General assessment
This manuscript presents LFS-SEnDA, a global ocean forecasting system incorporating a Sliding-Ensemble-based Data Assimilation (SEnDA) scheme. The study addresses a practically important problem in operational ocean forecasting and provides a relatively comprehensive description of the system configuration, assimilation methodology, computational implementation, and forecast evaluation.
The manuscript has several strengths. In particular, the use of a time-dependent historical ensemble provides a practical approach to representing temporal variations in background error statistics, while the optimized analysis implementation is relevant to the computational requirements of high-resolution global ocean forecasting. The evaluation results are generally encouraging and demonstrate the potential of the proposed system to improve forecast performance.
Overall, I consider this a valuable technical contribution. However, the manuscript could be improved by strengthening the connection between the technical design choices and their practical benefits. In particular, the presentation of the SEnDA methodology, the mathematical formulation and computational optimization, and the roles of several assimilation components would benefit from further clarification. My specific comments are as follows.
Major comments
1. Positioning and practical advantages of the sliding-ensemble approach
The manuscript presents the Sliding-Ensemble-based Data Assimilation scheme as one of its central components. However, the use of time-dependent or seasonally varying ensembles has already been explored in previous EnOI studies. Therefore, the methodological novelty of selecting historical ensemble members within a moving temporal window may be relatively limited in itself.
Rather than emphasizing this sampling strategy as a fundamentally new assimilation method, I suggest that the authors place greater emphasis on its practical value in the context of operational global ocean forecasting. For example, it would be useful to explain more clearly why this particular ensemble construction is suitable for the LFS system, how it accounts for temporal variations in background error statistics, and what practical advantages it offers in terms of implementation, computational cost, and forecast performance. The authors could also clarify whether the main contribution lies in the ensemble selection strategy itself or in its integration into an efficient and operationally applicable global forecasting workflow.
This would help establish a more balanced and convincing narrative, particularly given that the manuscript is primarily a technical system-development study rather than a proposal of a fundamentally new data assimilation theory.
2. Further evaluation of the computational performance and optimization
The manuscript provides a relatively clear explanation of the computational optimization, including the use of the Cholesky decomposition and the ensemble-space formulation. These descriptions help readers understand the rationale behind the proposed implementation. However, the performance evaluation appears to rely mainly on a single set of computational results, which may not be sufficient to demonstrate the overall efficiency and robustness of the optimized algorithm.
I suggest that the authors provide more comprehensive performance results, where feasible. For example, the computational cost could be evaluated under different ensemble sizes, observation numbers, or parallel configurations. It would also be helpful to report repeated-run statistics or other measures of performance variability, together with relevant information such as wall-clock time and parallel scalability. In addition, a clearer comparison between the original and optimized implementations would help quantify the actual benefits of the optimization.
The objective is not to require an extensive benchmarking study, but to provide sufficient evidence that the reported speedup is reproducible and that the optimized implementation remains efficient and stable under different computational conditions.
3. Clarification of the roles and contributions of the assimilation components
The SEnDA scheme incorporates several technical treatments, including ensemble spread adjustment, mixed-layer-related increment truncation, localization strategies, and the optimized analysis formulation. These components are potentially important for maintaining the physical consistency and computational efficiency of the assimilation system. However, their individual purposes and contributions to the final analysis and forecast performance are not always sufficiently clear.
I suggest that the authors strengthen the explanation of why these treatments are necessary and how they contribute to the overall assimilation framework. Where feasible, comparisons between configurations with and without selected components would provide useful evidence of their effectiveness. If additional experiments are not practical, a more explicit discussion of the underlying physical or numerical motivation, the specific problems each treatment is intended to address, and the expected consequences of omitting it would still improve the manuscript.
In particular, it would be helpful to distinguish the components that are essential to the assimilation methodology from those introduced primarily for numerical stability or computational efficiency. This would make the technical design of SEnDA easier to understand and reproduce.
Minor comments
1. Slight degradation in salinity forecast performance
Although the overall forecast performance is encouraging, the manuscript reports a slight degradation in salinity forecast skill in some depth ranges. This issue deserves some additional discussion. Salinity assimilation can be challenging in ensemble-based methods because of the limited ensemble spread, potentially weak or noisy temperature–salinity cross-covariances, and uncertainties in the background salinity field.
I encourage the authors to discuss possible reasons for this degradation and its relationship to the ensemble-based assimilation configuration used in this study. It would also be useful to mention potential technical improvements, such as more appropriate salinity spread adjustment, improved temperature–salinity cross-covariance representation, or variable-dependent localization and observation-error treatment. These approaches do not necessarily need to be implemented in the present study, but discussing them would provide a clearer perspective on the current limitations and possible future improvements.
2. Duration of the cycling experiment
The cycling experiment provides useful evidence of the stability of the proposed system, and the reported results appear generally consistent over the evaluation period. Nevertheless, the current experimental period of approximately 36 days is relatively short for assessing the stability of an operational ocean data assimilation and forecasting system.
If computational resources permit, I suggest extending the cycling experiment to at least two months. A longer period would provide stronger evidence of the temporal stability of the assimilation system and help determine whether the reported forecast improvements are sustained beyond the relatively short evaluation period. If such an extension is not feasible, the limitation should be acknowledged more explicitly.
Recommendation: Major Revision
The manuscript presents a technically valuable and practically relevant development of a global ocean forecasting system. The results are generally encouraging, and I do not consider the issues raised above to undermine the overall validity of the study. Most of my concerns relate to the clarity of the methodological positioning, the presentation of computational performance, and the explanation of the technical choices made in the assimilation system. Addressing these points would improve the manuscript's technical clarity and strengthen its overall narrative. I therefore recommend publication after major revision.