the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Dec 2023
Gap-filling processes on GOCI-derived daily sea surface salinity product for Changjiang diluted water front in the East China Sea
Abstract. The spatial and temporal resolutions of contemporary microwave-based sea surface salinity (SSS) measurements are insufficient. Thus, we developed gap-free gridded daily SSS product with high spatial and temporal resolutions, which can provide information on short-term variability in the East China Sea (ECS), such as the front changes by Changjiang diluted water (CDW). Specifically, we conducted gap-filling for daily SSS products based on the Geostationary Ocean Color Imager (GOCI) with a spatial resolution of 1 km (0.01°), using a machine learning approach during the summer season from 2015 to 2019. The comparison of the Soil Moisture Active Passive (SMAP), Copernicus Marine Environment Monitoring Service (CMEMS), and Hybrid Coordinate Ocean Model (HYCOM) SSS products with the GOCI-derived SSS over the entire SSS range showed that the SMAP SSS was highly consistent, whereas the HYCOM SSS was the least consistent. In the <31 psu range, the SMAP SSS was still the most consistent with the GOCI-derived SSS (R2 = 0.46; root mean squared error: RMSE = 2.41 psu); in the >31 psu range, the CMEMS and HYCOM SSS products showed similar levels of agreement with that of the SMAP SSS. We trained and tested three machine learning models—the find trees, boosted trees, and bagged trees models—using the daily GOCI-derived SSS as the ground truth, while including the three SSS products, environmental variables, and geographical data. We combined the three SSS products to construct input datasets for machine learning. Using the test dataset, the bagged trees model showed the best results (mean R2 = 0.98 and RMSE = 1.31 psu), and the models that used the SMAP SSS as input had the highest level. For the dataset in the >31 psu range, all models exhibited similarly reasonable performances (RMSE = 1.25–1.35 psu). The comparison with in situ SSS data, time series analysis, and the spatial SSS distribution derived from models showed that all models had proper CDW distributions with reasonable RMSE levels (0.91–1.56 psu). In addition, the CDW front derived from the model gap-free daily SSS product clearly demonstrated the daily oceanic mechanism during summer season in the ECS at a detailed spatial scale. Notably, the CDW front in the horizontal direction, as captured by the Ieodo Ocean Research Station (I-ORS), moved approximately 3.04 km per day in 2016, which is very fast compared with the cases in other years. Our model yielded a gap-free gridded daily SSS product with reasonable accuracy and enabled the successful recognition of daily SSS fronts at the 1-km level, which was previously not possible with ocean color data. Such successful application of machine learning models can further provide useful information on the long-term variation of daily SSS in the ECS.
- Preprint
(1060 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on essd-2023-421', Anonymous Referee #1, 25 Mar 2024
Dear Editor,
I have carefully reviewed the manuscript entitled Gap-filling processes on GOCI-derived daily sea surface salinity product for Changjiang diluted water front in the East China Sea. In this paper, the authors employ decision trees to derive a sea surface salinity (SSS) dataset for the Yellow Sea and East China Sea. Specifically, the input data comprises sea surface salinity (SSS) from three data products: GLORYS12, SMAP, and HYCOM, accompanied by other data such as sea surface temperature (SST), sea surface height (SSH), velocity, and wind stress curl. Notably, the "ground-truth" data used by the authors is the GOCI-derived SSS data from Kim et al. (2021), which was derived from SMAP data. The methods employed are three decision tree-based algorithms: fine trees, boosted trees, and bagged trees. The authors first compare the three SSS data products with the GOCI-derived SSS, then test different input combinations to determine the optimal inputs and algorithms. They demonstrate that the model with SMAP as input yields the highest accuracy, and the bagged trees perform best. However, models using the other two data products with bagged trees also produce satisfactory results. Comparisons between the SSS derived from the different models and SSS observations all yield coefficients of determination (R2) greater than 0.6. Ultimately, the authors decide to utilize the models with SMAP and CMEMS GLORYS data inputs as the "final results" for further analysis. In the last section, the authors use the SSS derived from Model 1 (SMAP input) and Model 2 (GLORYS input) to describe the evolution of the Changjiang Diluted Water in the East China Sea during the period of 2015-2019.
Despite the technical proficiency exhibited in applying fine trees, boosted trees, and bagged trees algorithms, and the methodological rigor in comparing and selecting data inputs and algorithms, I have significant reservations that compel me to recommend rejection of the manuscript for the following reasons:
- Ground-Truth Data Appropriateness: The designation of GOCI-derived SSS as "ground-truth" is fundamentally flawed. "ground-truth" refers to direct, in-situ measurements used to validate remote sensing products. GOCI-derived SSS from Kim et al. (2021), being a remotely sensed product itself, cannot serve as ground-truth, undermining the study's validation framework.
- Circular Logic in Methodology and Results: The manuscript's conclusion that SMAP SSS closely aligns with GOCI-derived SSS is tautological. Given that GOCI-derived SSS from Kim et al. (2021) was established using SMAP SSS data as "ground truth," it is logically unsound to use it as a benchmark for validation, leading to an inherent circularity in the comparative analysis.
- Presentation and Organization. Despite acknowledging the challenges faced by non-native English speakers, the manuscript's organization and language clarity fall below acceptable scientific standards. The presentation and organization of results, particularly in the manuscript's final section, are areas of concern. The narrative primarily enumerates outcomes from various models without effectively synthesizing these findings into coherent conclusions. This approach leaves readers struggling to discern the central thesis and implications of the research. It is critical for the authors to convincingly demonstrate the superiority and applicability of these products for future research with clear evidence. The manuscript currently falls short in this regard, relying on the reader’s inference rather than providing direct, substantiated arguments.
Additionally, it is particularly troublesome that crucial terms are consistently misspelled (e.g., "fine trees" as "find trees", even in the abstract), indicating a lack of thorough proofreading. While I respect the authors' efforts, the manuscript's frequent errors, especially in critical sections like the abstract, suggest a lack of thorough preparation. This oversight implies a disregard for the peer review process and makes me feel very disappointed. - Journal Scope Alignment. ESSD is a journal dedicated to "the publication of articles on original research data (sets), furthering the reuse of high-quality data of benefit to Earth system sciences". The primary emphasis of this manuscript on methodological comparison overshadows the utility and novelty of the data product itself, rendering it better suited for specialized remote sensing journals, such as IEEE Transactions on Geoscience and Remote Sensing, Remote Sensing, or Remote Sensing of Environment. Even for these journals, the first subsection comparing different data products would be more appropriate as supplementary information. In addition, the manuscript claims to present two final SSS products, one from SMAP (Model 1) and the other from GLORYS (Model 2). However, I checked the final output data via the link provided for data download (on ESSD website, the link in the manuscript is non-functional), only one SSS product is accessible. This discrepancy between the stated contributions and the actual accessible data undermines the research's completeness and poses significant concerns regarding data accessibility and transparency.
In summary, while the topic under investigation holds potential for advancing regional oceanographic research, the aforementioned concerns regarding methodology validation, manuscript quality, and alignment with journal scope are too substantive to overlook. A thorough revision addressing these fundamental issues is essential before reconsideration.
Despite the concerns highlighted, the achievement of an R^2 greater than 0.6 across various models is noteworthy and demonstrates the potential value of this study. From the perspective of a possible data user, I offer several suggestions aimed at enhancing the work's reliability and utility. These opinions are primarily based on two considerations: as a data user, how can I trust that this dataset is reliable? What details do I need from the author to utilize the data effectively? I highly recommend considering these improvements and look forward to the potential resubmission of this work.
General suggestions:
- Experiment Design and Data Utilization: In the development of remote sensing data products, two primary methodologies are commonly employed: (a) When observational samples (Y_obs) are scarce, an algorithm is designed to estimate values (Y_est) independently of these observations. The estimated values are then compared with the observed values (Y_obs) to evaluate accuracy. (b) With a large dataset of observations (Y_obs), the data is divided into two distinct subsets (Y1_obs and Y2_obs), collected at different times or during different cruises. The model is initially trained and tested with Y1_obs, followed by an independent evaluation using Y2_obs. Both strategies offer a convincing foundation for trusting the derived estimates (Y_est) in scenarios where direct observations are unavailable.
The authors opted for the latter strategy. However, the use of GOCI-derived SSS as "ground-truth" or observational data is problematic. I recommend that the authors consider the first approach in writing the manuscript, which does not require that the GOCI-derived SSS are “true values” or not. The significant contribution of this work should be the development of an algorithm capable of producing a SSS data product without relying on direct observations, validated independently through comparisons with NIFS data and (h–n) I-ORS data to ensure the product's reliability.
- Data Product Accessibility: provide a final data product. It is crucial to provide a clear and easily accessible final data product, instead of multiple choice from multiple models. Enhancing the accessibility of the derived SSS product would significantly improve the manuscript's utility to both readers and potential data users, making it a more practical resource in the field.
- Uncertainty Analysis. Offering an explicit estimation of the uncertainties associated with the final data product is essential for end-users. The authors should estimate uncertainties following methodologies from prior studies such as Wang et al., (2014) or Landschützer et al., (2014), thereby enhancing the data's reliability and user trust.
- Revisiting the Comparison Framework. The comparison between SMAP and GOCI-derived SSS may be more effectively presented after the final data product has been robustly defined and derived. A subsequent comparison of this final product with existing data products (SMAP, GLORYS, HYCOM) against independent observational data (from NIFS and I-ORS) would more convincingly demonstrate its advantages, establishing it as the current "best solution".
- Refinement of the Last Subsection. Simplify the discussion by focusing on the derived product's improved spatial or temporal resolution and its implications for studying environmental phenomena at new scales. Illustrating specific examples or case studies where the enhanced resolution provides novel insights would underscore the significance of this work in advancing our understanding of Earth system processes. Focus on showcasing advancements that were previously unattainable due to limitations in temporal and spatial precision. Make sure to clearly state these novel contributions in the opening or concluding sentence of the paragraph, ensuring that the reader immediately grasps the significance of the work's higher accuracy in both dimensions.
Ref:
Landschützer, P., Gruber, N., Bakker, D. C. E., & Schuster, U. (2014). Recent variability of the global ocean carbon sink. Global Biogeochemical Cycles, 28(9), 927–949. https://doi.org/10.1002/2014GB004853
Wang, G., Dai, M., Shen, S. S. P., Bai, Y., & Xu, Y. (2014). Quantifying uncertainty sources in the gridded data of sea surface CO 2 partial pressure. Journal of Geophysical Research: Oceans, 119(8), 5181–5189. https://doi.org/10.1002/2013JC009577Specific points:Figures and tables:1. Figure 2: “Ensemble classifier” in the bottom right corner of image should be “Ensemble regression”. Random Forest could be a classifier or regression learner, in this work it is obvious a regression learner.2. Figure 3: The colorbars for panels g to i are unclear. If these are 2-D density plots, the bin interval should be introduced.Main textIntroduction3. Line 40: “Because waters affected by river outflow and coastal regions are characterized by short- term variability, gridded SSS products can provide useful information for monitoring SSS variations”.This statement needs clarification. While gridded SSS products can indeed provide useful information for monitoring SSS variations, this capability is not necessarily due to the presence of river outflows or short-term variability in coastal regions.4. Line 57-59: “First, the accuracy of in situ observations defines how information is propagated from data-rich to data-sparse regions and is critically dependent on data coverage and the reliability of spatial covariance”.This statement needs revision. The abbreviated version of this sentence: "The accuracy of observations defined information propagation," and "The accuracy of observations depends on data coverage and spatial covariance"The accuracy of observations cannot define these. The author likely intends to convey that in regions with low observational coverage, the available data may not accurately represent the phenomena of interest.Data Section:5. Line 125: “The SMAP, HYCOM … were used …, HYCOM is a … model …. We used GOFS Global analysis data”.These sentences need to clarify the relationship between HYCOM and GOFS. It would be beneficial to reorganize the sentences to explicitly state the connection, such as: "HYCOM is a ... model, which forms the computational core of GOFS," or "HYCOM is a ... model within the GOFS."6. Line 129: “The GLORYS12V1 product is the CMEMS global ocean eddy-resolving reanalysis covering altimetry at 0.08° × 0.08° and 50 standard levels.”.However, the phrase "covering altimetry" may benefit from clarification. It is likely that "covering altimetry" implies either that the GLORYS12V1 product assimilates altimetry data into its reanalysis or that the GLORYS12V1 reanalysis spans the altimetry era, which began in 1993.7. Line 136: “at the highest spatial resolution”,The expression "highest" is inappropriate. It is recommended to change it to "at a high spatial resolution."8. Line 140: There is no introduction about the temporal resolution of ERA5. I guessed that the temporal resolution should be hourly.9. Line 153: “SSS data obtained from the ESC, West sea, and South sea”.The locations of "West sea and South sea" are unclear in Fig.1. Are the authors referring to the western and southern marginal seas of the Korean peninsula?10. Section 2.3. In situ data: An introduction regarding the uncertainty of the salinity measurements should be provided.Method section11. Line 172: “All data were sampled at 0.01°”,I guessed that the SSS, SSH, uo, vo, and wind stress curl were all interpolated into 0.01° resolution. Clarification on this point is requested.12. Line 181: “This method is more effective than other methods when the data vary rapidly”,This statement lacks objectivity, as it claims superiority over unspecified "other methods" without providing citations or clarifying the basis for comparison.13. Line 202: “the input groups … exhibited 500,000 matched pixel pairs or more”,It might be better to simply use "had" instead of "exhibited" for clarity.14. Line 203: “We then added as many as 10% of the matched pixel pairs to each input group.”.This sentence is unclear. The phrase "as many as" is unnecessary. Based on the following text, it seems the author intended to say, "For each training and test set, we then added 10% of its total number of zero matrices." Additionally, this approach is confusing, as the purpose and benefits are not introduced. A literature reference or justification for this method would be helpful, as it potentially reduces the signal-to-noise ratio.15. Line 208: “find trees”.The word “find” should be “fine”. The same mistake is also made in line 22 of the Abstract, Line 317 of the Section 4.2.1. Quantitative evaluation.16. Line 215 to 219: These sentences are not necessary.These sentences are not necessary, as the preceding sentences have already introduced the decision tree and two random forest ensemble algorithms. The reader can infer that the computational time will rank as follows: bagged trees > boosted trees > decision tree.17. Line 214. There is a typo; "booted trees" should be "boosted trees". To be honest I'm a bit disappointed, the authors shouldn't have made such a basic mistake in such important parts of the article.18. Section 3.2. It is confusing to use x to represent the GOCI-derived SSS (the "ground-truth" values) and y to represent the "compared SSS." It would be preferable to use y_tru to represent "ground-truth values" and y_est to represent "estimated values.".19. Equation (1) is the squared correlation coefficient, I prefer to use the equation: R^2 = 1 - [Σ(y_i - ŷ_i)^2 / Σ(y_i - ȳ)^2], as it directly represents the coefficient of determination, rather than being equivalent to the squared correlation coefficient. The authors can choose to deny my opinion on this point since they have the same values.Results and discussions20. Line 241: “distribution trends” is somewhat unusual and could be benefit from clarification or rephrasing. According to the following sentences, I think it should be “…, we examined the spatial and statistical distribution of xxx data products …”21. Line 258: “it did not reflect the daily SSS product because it was an 8-days average product”,The expression "reflect the daily product" is strange. It is guessed that the authors intended to convey that the SMAP dataset does not have a daily resolution.22. Line 296: “However, the SMAP SSS in our study area showed a more reasonable degree of agreement with the in situ NIFS SSS compared to that of the reanalysis SSS; hence, the SMAP SSS data can be a good reference for monitoring the CDW in the ECS.”.There are no figures or statistics provided in this manuscript to support this conclusion. Section 4.1 and Figure 3 are comparing three SSS data products with the GOCI-derived SSS.4.2 Performance of the SSS models23. Line 336: “Because to the difference in spatial resolution among the data, the pixels masked at zero were different;”,The expression "masked at zero" is misleading. I think the authors intended to convey that those pixels with missing values were different among the datasets. However, using "masked at zero" is misleading because temperature, salinity, and other variables have meaning when they equal zero.24. Line 338: “The model was trained using zero-free data, thereby not properly recognizing the actual mask pixels, resulting in a specific value of pixels in the masked area.”.The meaning of "specific value" in this context is unclear.25. Line 340: “we added as many as 10% of the matched pixel pairs for each model”,Again, clarification is needed regarding these 10% pixel pairs. Are they repeated samples using the bootstrap method or zero values?26. Line 381-385: Please provide the specific RMSE values. A 5.36% increase or 12.32% decrease may be small depending on the absolute value. For instance, assuming an RMSE of 1.5, those changes would only be 1.58 and 1.32, respectively.27. Line 393: “The RMSE of Model 5, which had the worst performance among the models.”,Grammatically, this is an incomplete sentence. It requires further elaboration.28. Line 394 – 396: “As shown in table 3, the three SSS datasets showed high accuracies in the >31 psu range; therefore, using the in situ NIFS dataset resulted in low RMSE values. However, the in situ I-ORS data have a high RMSE because most of the SSS data are in the <31 psu range.”NIFS and I-ORS are both observational datasets and, therefore, cannot have RMSE values themselves. It is the comparison between these observations and the three data products that yields the RMSE.29. Line 398: “In summary, all bagged trees models trained with a combination of various input variables could estimate the SSS of the CDW in the ECS on a daily basis with RMSE values of less than 1 psu, i.e., higher than that of the SMAP SSS.”There are two issues with this statement:a. The RMSE values of comparisons between NIFS and the models are larger than 1.0. Only the comparisons between most model-derived SSS and I-ORS have RMSE < 1 (i.e., Fig. 4h-n). Therefore, it should state "most bagged trees models" instead of "all models," as Figure 4j and 4l have RMSE = 1.0 and = 1.02, respectively.b. The phrase "higher than that of the SMAP SSS" should be replaced with "lower" or "better" to accurately convey the intended meaning.Data availability30. The link doesn’t work, and I go to the preprint dataset page, it should be https://doi.org/10.22808/DATA-2023-2Small typo:31. Line 64: “There are only few in situ SSS observations …”,“a few” or “few”? “few” means there are no in ARGO floats at all.32. Line 105: “because to” should be “because of”33. Line 180: “We applied to”, remove “to”34. Line 253: “.. all pixels in the study area cloud not provide SSS information”, It should be “could not” but “cloud not”.35. Line 286: “Currently, the SMAP are the only satellite data”, I think here “are” should be “is” because the subject SMAP is singular.36. Line 431: There is a typo; "patter" should be "pattern."37. Line 449: “We identified the three phases according the CDW”, “according” should be “according to”Citation: https://doi.org/10.5194/essd-2023-421-RC1 -
RC2: 'Comment on essd-2023-421', Anonymous Referee #2, 07 Apr 2024
Review of Gap-filling processes on GOCI-derived daily sea surface salinity product for Changjiang diluted water front in the East China Sea
by
Jisun Shin, Dae-Won Kim, So-Hyun Kim, Gi Seop Lee, Boo-Keun Kim, Young-Heon Jo
The paper presents a method to generate daily gap-filled sea surface salinity fields using as input lower resolution passive microwave data taking into account the correlation between ocean color and salinity fields and exploiting machine learning. The study focuses on the short term evolution of the Changjiang diluted water (CDW) front. It is essentially ok for publication after the recommendation below are followed.
My first recommendation is to modify the title that in its current form suggests that there are gap-filling processes at work in the East China Sea while I believe the authors should use instead of “Gap-filling processes on GOCI-derived…” the following: “Gap-filling techniques applied to GOCI-derived…”
It would important if the difference in methodology betweenKim, D. W., Kim, S. H., and Jo, Y. H.: Machine Learning to Identify Three Types of Oceanic Fronts Associated with the
Changjiang Diluted Water in the East China Sea between 1997 and 2021, Remote Sens., 14(15), 3574, doi:10.3390/rs14153574,
2022b and the current paper were clearly explained to show the reader the novelty of the current approach.
After the first definition of GOCI (Line 15) do not spell it out in the text and figure legends.
The following figures needs to be regenerated with increased resolution:
2, 3, 4, 5, 7 and 8
Detailed suggested changes to improve readability and correct apparent mistakes:
Line 13 Replace “high” with “higher”
Line 16 Replace “season from with “seasons of”
Line 17 Replace “Copernicus Marine Environment Monitoring Service” with “Copernicus Marine Service”
Line 30 Replace “horizontal” with “zonal”
Line 37 remove “—the salinity of the ocean at its surface—“
Line 28 replace “dataset” with “products”
Line 69 Replace “Copernicus Marine Environment Monitoring Service” with “Copernicus Marine Service”
LIne 130 replace and track with “along track”
Line 130-131 replace “using a reduced-order Kalman filter, and track altimeter data, satellite SST and in situ TS profiles
were jointly assimilated” with “using a reduced-order Kalman filter, along track altimeter data, satellite SST, and in situ TS profiles.”
Line 184 Replace “horizontal” with “zonal”
Line 431 Replace “patter” with “patterns”
Line 518 & 519 It seems from the manuscript that “>31 psu” (at line 518) should be swapped with “<31 psu” (at line 519)
Line 519 What is meant by “most oceanic environment”? “Typical” perhaps?
In Table 4: Model 1 (SMAP) with bagged trees reports the best RMSE 1.17 psu with 1.36 MSE however, we have either 1.36 MSE with 1.17 RMSE or 1.35 MSE with 1.16 RMSE. So a rounding off error was made. Since this model was selected as the best, from validation results, it is a relevant issue.
Citation: https://doi.org/10.5194/essd-2023-421-RC2
Data sets
gridded gap-free SSS dataset Jisun Shin, Dae-Won Kim, So-Hyun Kim, Boo-Keun Kim, Young-Heon Jo https://doi.org/10.22808/DATA-2023-2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 269 | 53 | 26 | 348 | 26 | 26 |
- HTML: 269
- PDF: 53
- XML: 26
- Total: 348
- BibTeX: 26
- EndNote: 26
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1