Reconstruction of High-Resolution Sea Surface Salinity over 2003–2020 in the South China Sea Using the Machine Learning Algorithm LightGBM Model

Round 1

Reviewer 1 Report

-The authors compare and combine different methodologies for the measurements of SSS: satellite, underway and station based. Add information about the sensors used, the accuracy of each method and explain the possible differences related to the “bulk” and the “skin” measurements from satellite.

-All the comparisons were done on seasonal averaged fields. Did the authors attempt a direct comparison between some station base observations, reconstructed field and concurrent SSS? This would help to quantify the differences for a better assessment of the performances of the method they developed and tested.

-Reconstructed fields fig.5 shows highest salinity on summer 2015 and 2016: can you explain why?

Underway SSS high values of winter salinity (fig.4) cannot be observed in the reconstructed field fig.5 Are those values below the standard deviations?

 -Comment the advantages and/or weakness of your method with respect to similar investigations performed by different authors.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Editor, Dear Authors

1. The main question addressed by the research

The authors reconstruct sea surface salinity in the South China Sea with a machine learning algorithm based on a combination of three datasets, and cruise observation-based dataset. The paper is based on the calibration (learning) of the algorithm and its validation using various statistical tests to assess the accuracy of the results. The approach is quite conventional; it has the merit of highlighting the surface salinity with various anomalies along the coast due to dilution of sea waters by fresh water issuing from rivers, and mixing with the Kuroshio current.

 

2. The topic is relevant in the field Work on the measurement of ocean surface salinity on a large scale is rare. They necessarily call upon complex techniques crossing various satellite or oceanographic measurements because high resolution spatial and temporal salinity data are not always available. In the present case the authors use recent techniques, i.e., machine learning algorithm based on a combination of remote sensing data and a large cruise observation-based dataset. This combination required a major effort in terms of the measurements carried out by the team concerning large cruise observation, which makes the study exhaustive

 

3. Compared to other published material the study presents a methodology leading to the high resolution spatial and temporal salinity measurement in the South China Sea. An exhaustive bibliographic study is made in the introduction.

 

4. Specific improvements the authors should consider is regarding the presentation of the results. According to me further controls could be considered to improve the measurement of salinity where the variations are important, i.e., within the Pearl River plume and the Kuroshio intrusion. But the main weakness of this work is the presentation of the results and their interpretation. The evolution of spatial distributions of sea surface salinity should be presented in a more condensed and more suggestive way. The work is carried out with rigor, the method is well explained but the authors must make a significant effort in terms of presentation of results. In particular, the figures 5, 6, 7, 9, 10, 11, 13 are unusable. Each of them could be replaced by a pair of maps representing the amplitude of the anomalies and their phase (the lag in relation to a specific date).

 

5. The conclusions are consistent with the evidence and arguments presented and they address the main question posed in the introduction.

 

6. The references are appropriate and up-to-date.

 

7. As already mentioned, the authors should focus their efforts on a better presentation of the results to clearly highlight the spatial and temporal resolution of the results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper reconstructed high-resolution sea surface salinity over 2003-2020 in the South China Sea with a machine learning algorithm. The topic of this paper fit within the stated scope of Remote Sensing and the paper is basically well-organized and well-written. However, several aspects could be further improved in order to having it published in this journal. The authors need to add some additional experiments to demonstrate the robustness of their methods. Therefore, my recommendation is major revision. The main questions I encountered when reading the manuscript are as follows:

 

1、Abstract is too long. Delete unimportant sentences and simplify it.

2、Line 46, delete ‘however’

3、Line 79, change ‘remote sensing derived’ to ‘satellite-derived’

4、Line 99, not accurate, satellite data is also used.

5、Line 135, It is necessary to validate your methods by separating underway data into two subset. However, the final products should use all underway data aw well as station-based observational SSS data. Namely, all available SSS data should be included in the LightGBM model to produce gridded SSS data. Furthermore, when selecting data for validation, you should conduct a series of experiments by randomly selecting a small amount of data. At present, the validation in not robustness enough.

6、Line 179, how testing subset is determined?

7、Line 187, LightGBM is widely-used in ocean science. May be some references can be given? Such as Gan et al. (2021), ‘Application of the Machine Learning LightGBM Model to the Prediction of the Water Levels of the Lower Columbia River’. DOI: 10.3390/jmse9050496

8、Figure2, can you show more details about how to generate gridded RS data from L3 data?

9、The observational SSS data used in this paper are publicly available? If not, I suggest the authors can upload them into the web. Also,you must provide the link of the reconstructed gridedd SSS data to let readers can freely obtain these data.

10、Line 524, all the authors contributed to the original writing? Really?

11、A suggestion, title can be revised as ‘Reconstruction of high-resolution sea surface salinity over 2003-2020 in the South China Sea using the maching learning LightGBM model’.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I think the paper can be published as is. Some effort has been made in the presentation.

Reviewer 3 Report

This paper can be accepted.