the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 16 May 2025
Reconstructing Sea Level Variability at the Ieodo Ocean Research Station (1993–2023) Using Artificial Intelligence, Machine Learning, and Reanalysis Integration
Abstract. This study presents a comprehensive approach for reconstructing a high-quality, continuous monthly sea level time series at the Ieodo Ocean Research Station (IORS) from 1993 to 2023 using advanced artificial intelligence (AI) and machine learning (ML) models. After applying quality control to the in-situ KIOST data, including inverse barometric effect correction, 3σ filtering, and a 75 % data coverage threshold, we validated trends using nearby PSMSL tide gauges and four ocean reanalysis datasets (CMEMS, GLORYS, ORAS5, HYCOM). The trend analysis showed a higher rate of sea level rise from in-situ data (4.94 mm/yr, Oct 2003–Dec 2023) compared to satellite and model-based estimates (e.g., CMEMS: 3.53 mm/yr, Jan 1993–Dec 2023), suggesting localized sea level rise in the East China Sea. Initial gap-filling used statistical models such as harmonic regression and regression-based climatology. A blended approach combining climatology and trend components achieved the best accuracy (RMSE ~0.056 m, R2 = 0.688). We then implemented various AI/ML models through an Iterative Imputer framework. Ensemble models (e.g., XGBoost) performed perfectly after 2003 but did not generalize well before 2004. Deep learning models like LSTM and GRU effectively captured seasonal and nonlinear patterns post-2003, with LSTM achieving RMSE = 0.023 m and R2 = 0.95. Time series models Prophet and SARIMA-SIN successfully reconstructed the full time series, with SARIMA-SIN estimating the highest trend (5.61 mm/yr). Multiple linear regression using reanalysis data served as a baseline, but AI/ML models outperformed it in both accuracy and generalization. This study provides a reproducible, interpretable, and physically consistent framework for reconstructing sea level variability in semi-enclosed coastal seas.
- Preprint
(2719 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on essd-2025-227', Anonymous Referee #1, 13 Jun 2025
The authors made an ambitious effort to reconstruct sea level variability from Ieodo Ocean Research Station using various models, AI/ML tools, and observations. I can appreciate the authors' diligent and thoughtful analysis. A few suggestions that I hope to make this manuscript better. First, throughout the manuscript, the KIOST in-situ data, KIOST sea level time series, IORS data, KIOST tide gauge, and IORS observation were used interchangeably. Do authors mean the same data set at IORS station maintained by the KIOST? Are there any other data sets from the KIOST? Please use the same acronym consistently to represent the data. Second, the main parameter compared is the linear “trend” from various models and observations, but I don’t see any regression plot on the manuscript to indicate how the trend is computed. Third, at the beginning of section 3, a significant upward trend of approximately 4.94 mm/yr from October 2003 to December 2023 and 5.43 mm/yr from January 2004 to December 2023 were reported. Is this from two different observation data sets? Why on the same location, only a few months of difference over a long 20 year period will have a different trend? Fourth, there are a few important figures that I really appreciate, i.e. Figure 5, 7 & 9. But too many lines with various legends get on top of each other. I cannot tell which one is which. For example, Figure 9 has 15 lines on top of each other. I suggest making the figures and legends larger and maybe make separated plots.
Other comments:
Line 16: What is KIOST? Acronym first appearance
Line 17: What is PSMSL? Never explained.
Line 21-22 “Initial gap-filling used statistical models such as harmonic regression and regression-based climatology”. → What are you trying to say?
Line 21-24: Unclear. Please rewrite.
Line 24-26 “Ensemble models (e.g., XGBoost) performed perfectly after 2003 but did not generalize well before 2004.” → This statement seems contradictory. Do you mean after 2003 is good, before 2004 is bad? How about 2003? Do you know why after 2003 is better? Can you speculate a reason? You don’t have observation prior the 2003, how do you know prior 2004 is bad?
Line 76 XGBooster? Acronym first appearance.
Line 81 IterativeImputer → Iterative Imputer
Line 101 Figure 1 is not mentioned in the text.
Line 101 Figure 1, What are KAN, LUS, SEO, FUR, NAK, NAS, NAH on the figure? I know they are on table 2, but ..
Line 113-114 Can you give a statistics/ percentage of how many data points were missing and data gaps?
Line 130: What is Epa, 2000?
Lien 137: Figure 2: What Good Quality Flag has data only @ ~2004 - 2008? Why do the “quality flags” have variations? What do you mean ‘flag’?
Line 163: Please explain the variables in Eqn (2).
Line 186: Please explain the variables in Eqn (4). What is Alfa_o Alfa_a?
Line 349 IterativeImputer → Iterative Imputer
Line 370-372 “a significant upward trend of approximately 4.94 mm/yr from October 2003 to December 2023 and 5.43 mm/yr from January 2004 to December 2023.” → Can you explain how these two numbers of the upward trend were obtained. Did you make a regression fit? Can you show them in a figure? Why is it only a few months different (Oct 2003 vs Jan 2004) over a long period of 20 years that the upward trend has so much difference? If you want to compare the upward trend of IORS and other models, why don’t you use the same period of time?
Line 375 Table 2, caption. “five PSML tide gauge stations “ → There are seven stations listed on Table 2.
Line 475 Figure 7: This is a great figure which shows some AI/ML models can reproduce the IORS observations. Unfortunately, there are too many lines, I can not tell which one is which. Can you either reduce the number of lines or make separated plots?
Line 669- “Among AI/ML models, ensemble learners (e.g., 670 XGBoost, RandomForest) achieved perfect reconstruction metrics after September 2003 but failed to predict values in earlier periods”. → You don’t have observation data prior to 2003, how do you know “failed to predict values in earlier periods”?
Citation: https://doi.org/10.5194/essd-2025-227-RC1 -
AC1: 'Reply on RC1', MyeongHee Max Han, 14 Jun 2025
The authors made an ambitious effort to reconstruct sea level variability from Ieodo Ocean Research Station using various models, AI/ML tools, and observations. I can appreciate the authors' diligent and thoughtful analysis. A few suggestions that I hope to make this manuscript better.
- We sincerely appreciate your careful and constructive comments, which have greatly helped us improve the clarity and quality of the manuscript. I have addressed all of your questions and suggestions based on the revised manuscript. Unfortunately, I am unable to upload the revised manuscript directly here, but I would be happy to provide further clarification or additional information if needed. Please do not hesitate to let me know if you have any additional questions or comments.
First, throughout the manuscript, the KIOST in-situ data, KIOST sea level time series, IORS data, KIOST tide gauge, and IORS observation were used interchangeably. Do authors mean the same data set at IORS station maintained by the KIOST? Are there any other data sets from the KIOST? Please use the same acronym consistently to represent the data.
- Thank you for pointing this out. We confirm that all these terms refer to the same in-situ sea level dataset recorded at the Ieodo Ocean Research Station (IORS), maintained by the Korea Institute of Ocean Science and Technology (KIOST). In the revised manuscript, we now consistently use “IOSR” to refer to this dataset throughout.
Second, the main parameter compared is the linear “trend” from various models and observations, but I don’t see any regression plot on the manuscript to indicate how the trend is computed.
- We appreciate this important suggestion. To clarify, linear trends were computed using ordinary least squares regression of monthly sea level anomalies against time (in months). To improve transparency, we have added a new figure (Figure 8) showing both the monthly time series and their corresponding regression lines for IORS (with trends over 2003–2023 and 2004–2023), CMEMS, GLORYS, ORAS5 (all over 1993–2023), and HYCOM (over its valid range, 1994–2023). This revised figure and caption explain the specific time windows used for each regression.
Third, at the beginning of section 3, a significant upward trend of approximately 4.94 mm/yr from October 2003 to December 2023 and 5.43 mm/yr from January 2004 to December 2023 were reported. Is this from two different observation data sets? Why on the same location, only a few months of difference over a long 20 year period will have a different trend?
- Both trends were computed from the same IORS dataset. The 4.94 mm/yr trend corresponds to October 2003 to December 2023, beginning with the first available observation, while the 5.43 mm/yr trend covers January 2004 to December 2023, aligning with full calendar years. The difference (~0.5 mm/yr) reflects the sensitivity of linear regression to initial values, especially relatively high sea levels in October and November 2023, which exert downward leverage on the trend estimate. As a result, the earlier start yields a slightly lower (less steep) slope. This explanation has been incorporated into the manuscript, and both regression lines are shown in Figure 8 to illustrate the impact of the starting point. To ensure consistent comparison across datasets, Table 5 has been revised to report trends over the common period January 2004 to December 2023 for all satellite and reanalysis products (CMEMS, GLORYS, ORAS5, HYCOM), while also retaining the full-period trends (1993–2023 or 1994–2023) for completeness.
Fourth, there are a few important figures that I really appreciate, i.e. Figure 5, 7 & 9. But too many lines with various legends get on top of each other. I cannot tell which one is which. For example, Figure 9 has 15 lines on top of each other. I suggest making the figures and legends larger and maybe make separated plots.
Thank you for this helpful suggestion. We revised the figures as follows:
- Figure 5: We introduced vertical offsets and reordered the legend by distance to the IORS.
- Figure 7: We adopted a similar offset style and added trend values to the legend.
- Figure 9: We revised the figure by applying vertical offsets to the reconstructed sea level time series from each model to reduce visual overlap. Models are labeled with both offset and linear trend values in the legend
These changes significantly improve figure readability, and we appreciate your guidance.
Other comments:
Line 16: What is KIOST? Acronym first appearance
- We clarified the first mention of KIOST as “Korea Institute of Ocean Science and Technology (KIOST)” in Line 112 and removed the redundant mention in Line 16.
Line 17: What is PSMSL? Never explained.
- PSMSL is now defined as “Permanent Service for Mean Sea Level (PSMSL)” in Lines 17–18.
Line 21-22 “Initial gap-filling used statistical models such as harmonic regression and regression-based climatology”. → What are you trying to say?
- We have rewritten the sentence as “Initial gap-filling of the in-situ IORS sea level data was conducted using statistical models, including harmonic regression and regression-based climatology.” in Lines 22-23.
Line 21-24: Unclear. Please rewrite.
- We have rewritten the sentence as “Initial gap-filling of the in-situ IORS sea level data was conducted using statistical models, including harmonic regression and regression-based climatology. A blended approach, integrating climatological cycles with a linear trend, yielded the highest accuracy when validated against observed (non-missing) data (RMSE ≈ 0.056 m; R² = 0.688).” in Lines 22-25.
Line 24-26 “Ensemble models (e.g., XGBoost) performed perfectly after 2003 but did not generalize well before 2004.” → This statement seems contradictory. Do you mean after 2003 is good, before 2004 is bad? How about 2003? Do you know why after 2003 is better? Can you speculate a reason? You don’t have observation prior the 2003, how do you know prior 2004 is bad?
- We have rewritten the sentence as “We then implemented various AI/ML models through an Iterative Imputer framework. Ensemble models (e.g., XGBoost) accurately reproduced IORS observations after October 2003 but generated unrealistic values before this period, likely due to overfitting and limited ability to extrapolate beyond the training data range.” in Lines 25-29.
Line 76 XGBooster? Acronym first appearance.
- We have corrected to “Extreme Gradient Boosting (XGBoost).” in Line 80.
Line 81 IterativeImputer → Iterative Imputer
- We have corrected to “Iterative Imputer” in Line 85 and the main text; we retain IterativeImputer as the code reference in the Methods section.
Line 101 Figure 1 is not mentioned in the text.
- We now cite Figure 1 in the main text in Line 45.
Line 101 Figure 1, What are KAN, LUS, SEO, FUR, NAK, NAS, NAH on the figure? I know they are on table 2, but ..
- The figure caption now includes “KAN, LUS, SEO, FUK, NAK, SIM, NAS, and NAH represent KANMEN, LUSI, SEOGWIPO, FUKUE, NAKANO, SIMA, NASE III, and NAHA” in the caption of Figure 1 in Line 105.
Line 113-114 Can you give a statistics/ percentage of how many data points were missing and data gaps?
- We have added “Of the 372 expected monthly records from 1993 to 2023, 189 entries (50.8%) were missing. From the start of continuous observations in October 2003 to December 2023, 60 out of 243 months (24.7%) contained missing data.” In Lines 121-123.
Line 130: What is Epa, 2000?
- We have corrected to “(U.S. Environmental Protection Agency, 2000)” in Lines 139-140.
Lien 137: Figure 2: What Good Quality Flag has data only @ ~2004 - 2008? Why do the “quality flags” have variations? What do you mean ‘flag’?
- We have revised the caption to explain that gray dots represent 10-minute data flagged as "good quality." Their prominence around 2004–2008 reflects changes in instrumentation and overplotting in Lines 147-157.
Line 163: Please explain the variables in Eqn (2).
- We have explained the variables in Lines 181-186.
Line 186: Please explain the variables in Eqn (4). What is Alfa_o Alfa_a?
- We have explained the variables including and in Lines 209-211.
Line 349 IterativeImputer → Iterative Imputer
- We have corrected to Iterative Imputer in Line 373.
Line 370-372 “a significant upward trend of approximately 4.94 mm/yr from October 2003 to December 2023 and 5.43 mm/yr from January 2004 to December 2023.” → Can you explain how these two numbers of the upward trend were obtained. Did you make a regression fit? Can you show them in a figure? Why is it only a few months different (Oct 2003 vs Jan 2004) over a long period of 20 years that the upward trend has so much difference? If you want to compare the upward trend of IORS and other models, why don’t you use the same period of time?
- The trends were derived using ordinary least squares linear regression and are illustrated in Figure 8. Relatively high sea level values in October and November 2003 exert downward leverage on the regression line, resulting in a slightly lower trend estimate when this period is included (Lines 395-399). To ensure consistency across datasets, all trend comparisons in Table 5 use the period January 2004 to December 2023. Table 5 has also been added to explicitly demonstrate the influence of the initial months on the calculated trend values.
Line 375 Table 2, caption. “five PSML tide gauge stations “ → There are seven stations listed on Table 2.
- We have corrected Line 403 in Table 2 caption to accurately indicate seven PSMSL tide gauge stations are included.
Line 475 Figure 7: This is a great figure which shows some AI/ML models can reproduce the IORS observations. Unfortunately, there are too many lines, I can not tell which one is which. Can you either reduce the number of lines or make separated plots?
- Figure 7 was redrawn using separated offset plots, improving clarity while preserving the information in Line 508.
Line 669- “Among AI/ML models, ensemble learners (e.g., 670 XGBoost, RandomForest) achieved perfect reconstruction metrics after September 2003 but failed to predict values in earlier periods”. → You don’t have observation data prior to 2003, how do you know “failed to predict values in earlier periods”?
- We acknowledge that no observational data exist before October 2003. Our comment about ensemble models “failing to predict” earlier values refers to their unrealistic and flat outputs in this period, as seen in Figure 7. This was not a validation failure but a lack of generalization in the absence of physical context. When supplemented with satellite/reanalysis predictors (Figure 9), these models performed well across the entire period (1993–2023).
-
AC1: 'Reply on RC1', MyeongHee Max Han, 14 Jun 2025
-
RC2: 'Comment on essd-2025-227', Anonymous Referee #2, 27 Jul 2025
The paper by Han and Lim provides an overview of methods to reconstruct
monthly sea level values for a measuring station in an enclosed sea.I believe a number of issues with the current version of the paper makes
it unsuitable for publication in ESSD, in particular:
- the level of information on the input datasets is insufficient, as well
as their overall referencing (e.g. no references to models in Section 2.3!)
- multiple ML models are used to perform data imputation, but the model
setup is unclear and how observation/model data are discriminated in the
files is not mentioned
- Related to that, it is unclear how train/test split were prepared. The
"perfect" performances of some models might be indicative of data leakage
(as presented in Table 4), so this should be explained. No information on model hyperparameters is
presented (e.g., losses, epochs...), hampering reproducibility
- Using both "AI" and "ML" seems unnecessary, unless I am missing something
only ML suffices
- The current use of reanalyses is puzzling. While they may be used to learn
sea level forcings and bias correct them when IORS data are available, and then
reconstruct observation (withdrawing reanalysis sea level!) in 1993-2003.
Such reconstruction should be carefully validated if independent data exist.
A map with mean sea level and trends from reanalyses as Fig. 1 could be useful to include,
given the large gaps in Fig.8
- I am missing what makes IORS unique, if similar datasets exist in the region,
and key information (Its coordinates? Is it an island or a mooring in the open ocean?
Which instruments are there? How many observations are missing?). Photos and relevant details are needed.
A dedicated subsection is required in section 2.
L1 mentioning East China Sea would clarify scope
L2 using both AI and ML is misleading
L13-32 too many unspecified acronyms (e.g. KIOST, CMEMS...)
L25 This sounds unlikely. A "perfect" performance sounds like data leakage to me
L55 these two references seems unsuitable; please find more general reviews or
textbooks to cite here
L61 "dense" is unclear; you mean spatially?
L69 what do you mean with dummy variables?
L71 why "realistic"?
L79 these are deep networks
L83 I would say this is from Meta https://facebook.github.io/prophet/
L90 The Copernicus Marine Service (not CMEMS!) provides different reanalyses:
https://data.marine.copernicus.eu/products?facets=areas%7EGlobal+Ocean--sources%7ENumerical+models--mainVariables%7ESea+surface+height
Please clarify exactly what you are using
L94 in-situ data can be used to validate models, not the other way round! rephrase
L96 These products cannot be considered auxiliary, since the same variable is used
and presumably data is assimilated in the area
L112 Is Kiost an author?
L113 "Intermittently" -> provide numbers
L119 where are pressure measurements taken from?
L125 This levelof detail is insufficient. Instrumentation information and algorithmic
details must be provided or fully referenced (what is a "good diagnostic"?)
L127 how is the standard deviation computed? Across all measurements in one or multiple
months?
L132 Why only giving monthly means if daily data are accessible?
L134 with "foundation" you mean "input"?
Fig 2 why the grey time series is so short, and why using range 1993-2023 if data is only
post 2003? The figure seems redundant with Fig. 3, while a table with number of observations,
gaps and periods should be added
L164 I guess only linear long-term
L167 What's this acronym?
L190 Why "realistic"? Maybe "empirical" or "combined" is a better word?
L195 a very specific reference, is there a more general to mention?
L216-7 one or two studies? This is confusing
L236 "widely" and then one study...does not seem so wide
L246 This is confusing. Which code are you using? catboost should be open source https://catboost.ai/docs/en/concepts/python-installation
L220 a verb is missing here
L266 ARIMA undefined and unreferenced
L273 unsupported, add reference
L279 I don't see how holidays matter for sea level
L296 Please clarify overfitting in your training strategy
L307 what does "D" mean?
L324 what is y_{ll}?
L338 there are no missing values in reanalysis, unless over land. Clarify.
Also, is HYCOM a reanalysis? Add references for all products here
L352 This is not multivariate if only sea level is used, revise
L361 How is this physically informed? Via the loss?
Table 1 the resolution here are not the native ones?
Table 2 what's PSMSL?
L384 this 5.8 m is unexplained
L390 isn't the following list redundant with the Methods section?
Table 3 what do you mean with "forced"?
L431 how this is performed is unclear. Training and validation periods should be fully disjoint. Please explain
L444 This is puzzling, I wonder if they don't work well since they would best work on chunks,
unlike e.g. Prophet. Please clarify
Table 6 why two periods in the first row?
L487 Was IORS already explained?
L493 the statement on assimilation is key: if reanalyses assimilated in situ data, they are
not independent, so they cannot be validated then. The spatial resolution of reanalyses should be given
Fig 9 and similar are not readable. Validation periods should be shown. Moreover, with the
current setup I have the impression just a combination of reanalysis will explain 1993-2002
L587 This section seems misplaced
L646 tuning not detailed elsewhere
L688 I believe access to the dataset should be provided, and full information should be given
in the files (e.g., to distinguish observational and model-based data)
The naming here https://www.seanoe.org/data/00865/97666/ is unclearCitation: https://doi.org/10.5194/essd-2025-227-RC2 -
AC2: 'Reply on RC2', MyeongHee Max Han, 01 Aug 2025
Thank you very much for your thoughtful and detailed comments.
I am currently on a business trip this week and may not be able to respond to all of your points immediately. However, I truly appreciate your valuable feedback—it has been very helpful in guiding my revisions.
I will carefully address your comments and provide a revised manuscript as soon as possible.
Thank you again for your insights and support.
Citation: https://doi.org/10.5194/essd-2025-227-AC2
-
AC2: 'Reply on RC2', MyeongHee Max Han, 01 Aug 2025
Data sets
KIOST Combined SeaLevels at the IORS MyeongHee Han et al. https://doi.org/10.17882/97666
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 270 | 43 | 19 | 332 | 12 | 23 |
- HTML: 270
- PDF: 43
- XML: 19
- Total: 332
- BibTeX: 12
- EndNote: 23
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
