Next Article in Journal
Numerical Simulation of Thermo-Hydro-Mechanical Coupling of Model Test for Nuclear Waste Disposal
Next Article in Special Issue
Using Eye-Tracking Data to Examine Response Processes in Digital Competence Assessment for Validation Purposes
Previous Article in Journal
Laser Applications in Metal Orthodontic Bracket Debonding: A Systematic Review
Previous Article in Special Issue
TopoSinGAN: Learning a Topology-Aware Generative Model from a Single Image
 
 
Article
Peer-Review Record

Monitoring Sea Surface Temperature and Sea Surface Salinity Around the Maltese Islands Using Sentinel-2 Imagery and the Random Forest Algorithm

Appl. Sci. 2025, 15(2), 929; https://doi.org/10.3390/app15020929
by Gareth Craig Darmanin, Adam Gauci, Monica Giona Bucci and Alan Deidun *
Reviewer 1:
Reviewer 2: Anonymous
Reviewer 3: Anonymous
Appl. Sci. 2025, 15(2), 929; https://doi.org/10.3390/app15020929
Submission received: 18 December 2024 / Revised: 14 January 2025 / Accepted: 16 January 2025 / Published: 18 January 2025
(This article belongs to the Special Issue Advances and Applications of Complex Data Analysis and Computing)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The title of the paper seems somewhat awkward and inappropriate. I recommend revising it as follows:

Original: Satellite Remote Sensing and Machine Learning for Sea Surface Temperature and Sea Surface Salinity Monitoring around the Maltese Islands

Revised: Sea Surface Temperature/Salinity Monitoring Using Sentinel-2 Imagery Around the Maltese Islands Based on the Random Forest Algorithm

A comprehensive workflow summarizing the application of the Random Forest (RF) algorithm, including the input and validation data, should be presented as an additional figure.

The longitude and latitude coordinates of the study area should be added to Figures 3, 4, and 5.

Since Table 2 contains identical content except for the year, it would be better to explain this in the main text rather than presenting it as a table.

The references or citations are not the latest. Including recent research studies to enhance the paper's relevance is necessary.

This paper does not propose a new processing method or technique but presents a case study of applying an existing algorithm. Therefore, the input data and validation methods used must be explicitly presented.

The explanation that the lower PCC and RMSE in 2024 compared to 2022 and 2023 are due to the difference in the number of training data (40 vs. 400) seems questionable. This interpretation should be critically evaluated.

The use of model data instead of field data for model validation might affect the reliability of the results. What are the possible solutions to address this issue?

In Figure 8, the legend values for SSS and SST are different, making it impossible to observe the annual variation. Specifically, SSS 2022 (37.69-37.79), SSS 2023 (37.55-38.55), and SSS 2024 (37.20-38.20). The legends should be unified for clarity and redraw the figure.

Line 290: “nrr estimators” Is this correct? Please clarify.

Lines 413-425: This study processes SSS and SST separately and provides qualitative interpretations of their correlations. Instead of limiting the analysis to textual interpretations, quantitative analysis is needed.

Author Response

Comment 1: The title of the paper seems somewhat awkward and inappropriate. I recommend revising it.

Response 1: Monitoring Sea Surface Temperature and Sea Surface Salinity Around the Maltese Islands Using Sentinel-2 Imagery and the Random Forest Algorithm.

Comment 2: A comprehensive workflow summarizing the application of the Random Forest (RF) algorithm, including the input and validation data, should be presented as an additional figure.

Response 2: Added Figure 7 showing the workflow used to create the prediction model.

Comment 3: The longitude and latitude coordinates of the study area should be added to Figures 3, 4, and 5.

Response 3: The longitude and latitude coordinates have been added to Figures 3, 4, and 5.

Comment 4: Since Table 2 contains identical content except for the year, it would be better to explain this in the main text rather than presenting it as a table.

Response 4: This is explained within the text in Sections 2.2 and 2.3, the table is implemented to provide a summary of these two sections making it easier for the reader to understand.

Comment 5: The references or citations are not the latest. Including recent research studies to enhance the paper's relevance is necessary.

Response 5: I have added more recent references (2, 3, 4, 34, 35, 36) in the introduction and section 3.2 of the results.

Comment 6: The input data and validation methods used must be explicitly presented.

Response 6: Added a paragraph in section 2.5 describing the input data and validation methods used between line 351-364.

Comment 7: The explanation that the lower PCC and RMSE in 2024 compared to 2022 and 2023 are due to the difference in the number of training data (40 vs. 400) seems questionable. This interpretation should be critically evaluated.

Response 7: This has been better explained between line 424-434 in Section 3.1.

Comment 8: The use of model data instead of field data for model validation might affect the reliability of the results. What are the possible solutions to address this issue?

Response 8: Possible solution to mitigate the use of model data is described in the conclusion between line 593-599.

Comment 9: In Figure 8, the legend values for SSS and SST are different, making it impossible to observe the annual variation. Specifically, SSS 2022 (37.69-37.79), SSS 2023 (37.55-38.55), and SSS 2024 (37.20-38.20). The legends should be unified for clarity and redraw the figure.

Response 9: Initially, this was the intended approach; however, after standardising the legend, two of the maps appeared blank because the legend values varied between the maps.

Comment 10: Line 290: “nrr estimators” – Is this correct? Please clarify.

Response 10: This was changed to ‘ntree’ signifying the number of trees.

Comment 11: Lines 413-425: This study processes SSS and SST separately and provides qualitative interpretations of their correlations. Instead of limiting the analysis to textual interpretations, quantitative analysis is needed.

Response 11: I have included references to support my argument. However, while I appreciate the suggestion, I do not believe that quantitative analysis is necessary for this section, as the focus of this paper is not on examining the relationship between SSS and SST, but rather on assessing the accuracy of the Random Forest model in predicting the parameters.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents a study that leverages satellite remote sensing and machine learning to monitor SST and SSS around the Maltese Islands. The integration of Sentinel-2 satellite data with in-situ measurements and the application of the RF algorithm represent a methodological approach. The study's findings could contribute insights to marine environmental monitoring and climate change studies in the region. However, several aspects of the manuscript require further refinement and clarification to enhance its rigor. 

Major Comments:

The primary concern lies in the generalization and transferability of the results. While the study demonstrates strong performance of the RF algorithm in predicting SST and SSS, it is crucial to understand how these results may vary under different environmental conditions or in other geographic locations. The manuscript would benefit from a more thorough discussion on the external validity of the model.

In Section 3.2, the manuscript overuses speculative language such as "may" and "might" without sufficient literature support or quantitative analysis to back up the claims. The discussion lacks the depth needed to solidify the findings, and the impact factors are not adequately explored. To improve this section, the authors should incorporate more concrete evidence from existing research or consider including quantitative analyses to substantiate the suggested relationships between environmental changes and the variations in SSS and SST. This will lend greater credibility and rigor to the study's conclusions.

The paper mentioned the selection of the RF algorithm, but did not compare it with other potential algorithms. A brief discussion on why RF was chosen over other machine learning models, such as neural networks or support vector machines, would enhance the persuasiveness of the research methods. More importantly, it is necessary to clarify by how much RF outperforms traditional methods or existing inversion methods.

Minor Comments:

Figure 1 lacks information on longitude and latitude.

In Figure 2, why are the "13 Spectral Bands from Sentinel-2, 2021 image" placed in the "In-situ Dataset Fields"?

In Figures 3 and 4, please use different styles of dots to distinguish the data used for training and testing.

Author Response

Comment 1: The primary concern lies in the generalization and transferability of the results. While the study demonstrates strong performance of the RF algorithm in predicting SST and SSS, it is crucial to understand how these results may vary under different environmental conditions or in other geographic locations. The manuscript would benefit from a more thorough discussion on the external validity of the model.

Response 1: This paper primarily focuses on SST and SSS, and as such, other variables have not been examined to determine whether this methodology can be applied to them. Nevertheless, a section in the conclusion has been included that outlines potential future work, one aspect of which involves extending this workflow to predict other marine parameters (Lines 585-602).

Comment 2: In Section 3.2, the manuscript overuses speculative language such as "may" and "might" without sufficient literature support or quantitative analysis to back up the claims. The discussion lacks the depth needed to solidify the findings, and the impact factors are not adequately explored. To improve this section, the authors should incorporate more concrete evidence from existing research or consider including quantitative analyses to substantiate the suggested relationships between environmental changes and the variations in SSS and SST. This will lend greater credibility and rigor to the study's conclusions.

Response 2: I have included references to support my argument. However, while I appreciate the suggestion, I do not believe that quantitative analysis is necessary for this section, as the focus of this paper is not on examining the relationship between SSS and SST, but rather on assessing the accuracy of the Random Forest model in predicting the parameters.

Comment 3: The paper mentioned the selection of the RF algorithm, but did not compare it with other potential algorithms. A brief discussion on why RF was chosen over other machine learning models, such as neural networks or support vector machines, would enhance the persuasiveness of the research methods. More importantly, it is necessary to clarify by how much RF outperforms traditional methods or existing inversion methods.

Response 3: I have included a paragraph in Section 2.5 explaining the rationale behind my choice of the RF algorithm and comparing it with other potential alternatives (Lines 301-312). However, I did not specify the extent to which the RF outperforms traditional methods, as this can vary depending on the specific case and is not universally fixed.#

Comment 4: Figure 1 lacks information on longitude and latitude.

Response 4: Longitude and latitude information has been added

Comment 5: In Figure 2, why are the "13 Spectral Bands from Sentinel-2, 2021 image" placed in the "In-situ Dataset Fields"?

Response 5: This is incorporated in this field as the in-situ dataset comprises both the data points collected from the seaglider missions and the Sentinel-2 band data corresponding to each individual in-situ point.

Comment 6: In Figures 3 and 4, please use different styles of dots to distinguish the data used for training and testing.

Response 6: This is not feasible, as the algorithm randomly selects the training and testing data points, meaning that with each iteration, the chosen points vary.

Reviewer 3 Report

Comments and Suggestions for Authors

In the article authors present an innovative approach to monitoring spatial and temporal variability in sea surface temperature (SST) and salinity (SSS) around Malta, utilizing remote sensing technologies, in-situ data, and the Random Forest (RF) algorithm. The authors demonstrate that integrating these methods allows for precise parametric mapping, contributing to a better understanding of oceanographic processes. The results highlight the effectiveness of the applied techniques in modeling dynamic oceanic parameters, though limitations related to data size and quality significantly affect prediction accuracy. The article provides an important contribution to the development of marine environment monitoring methods while emphasizing the need for further advancements in technology and methodology in this field. Below are some comments and suggestions for authors:

The introduction effectively outlines the importance of monitoring SSS and SST in the context of global climate changes and the specific characteristics of the Mediterranean environment. However it could provide more detailed discussion of the challenges associated with monitoring marine parameters around Malta, such as the impact of local currents or diverse atmospheric conditions. Include more recent studies published in 2023–2024 that reflect advancements in machine learning applications for marine monitoring (e.g., neural networks, hybrid models) and specific challenges in integrating in-situ and satellite data for predicting dynamic marine parameters. Highlighting research gaps in the application of RF algorithms in this field would also enhance the context.

The methodology is detailed, including descriptions of in-situ data collection, satellite data utilization, and the application of the RF algorithm but  details regarding the optimization of RF hyperparameters are too general. Providing more information about the tuning process and analysis of how various parameters impact results would improve transparency and reproducibility.

The results are presented in an accessible manner, and the visualizations (tables, dot plots, spatial maps) are clear and appropriate. Noteworthy is the high correlation between actual and projected data for 2022-2023.I suggest a more detailed discussion of the impact of dataset size limitations (especially for 2024) on the decline in correlation rates. In addition, the discrepancies in the results for SSS and SST, particularly the higher variability of the SST data, require further clarification. Are these methodological differences or inherent in the data? Additionally, enhancing the maps with statistical analyses (e.g., regions with the highest variances) would enrich the presentation of the results.

The conclusions are too general and do not fully address the impact of methodological limitations on the results. The need for larger data sets and improved models should be emphasized. I suggest a more critical approach to the results, including specific recommendations for methodological improvements to improve prediction accuracy.

To summarize, the article presents significant achievements in applying remote sensing and machine learning methods to monitor oceanographic parameters. While the methodology is detailed and the results promising, limitations related to data quality and size restrict the generalization of findings. I recommend accepting the article after implementing the following revisions:

  • expand the introduction to include specific local challenges and supplement the literature with studies specific to the Mediterranean Sea
  • provide a more detailed description of RF algorithm optimization and justify the selection of variables
  • critically analyze the results, addressing limitations related to data
  • enhance the conclusions with more detailed proposals for future research and solutions to methodological issues

The article makes an important contribution to the field, especially regarding studies of dynamic marine parameters, but requires revisions to fully acknowledge its limitations.

   

 

 

Author Response

Comment 1: The introduction effectively outlines the importance of monitoring SSS and SST in the context of global climate changes and the specific characteristics of the Mediterranean environment. However it could provide more detailed discussion of the challenges associated with monitoring marine parameters around Malta, such as the impact of local currents or diverse atmospheric conditions. Include more recent studies published in 2023–2024 that reflect advancements in machine learning applications for marine monitoring (e.g., neural networks, hybrid models) and specific challenges in integrating in-situ and satellite data for predicting dynamic marine parameters. Highlighting research gaps in the application of RF algorithms in this field would also enhance the context.

Response 1: I have expanded my introduction to provide a more detailed explanation of the importance of understanding these marine parameters, particularly in the context of the Mediterranean Sea. Additionally, I have incorporated recent studies to further support and strengthen the introduction. Furthermore, in section 3.3, I have included a paragraph discussing the limitations of using satellite and in-situ data to predict highly dynamic parameters.

Comment 2: The methodology is detailed, including descriptions of in-situ data collection, satellite data utilization, and the application of the RF algorithm but  details regarding the optimization of RF hyperparameters are too general. Providing more information about the tuning process and analysis of how various parameters impact results would improve transparency and reproducibility.

Response 2: A paragraph has been included in section 2.5 outlining the process and steps involved in generating the RF model (Lines 351-364), accompanied by a schematic (Figure 7) to facilitate the reader's understanding of the process.

Comment 3: The results are presented in an accessible manner, and the visualizations (tables, dot plots, spatial maps) are clear and appropriate. Noteworthy is the high correlation between actual and projected data for 2022-2023.I suggest a more detailed discussion of the impact of dataset size limitations (especially for 2024) on the decline in correlation rates. In addition, the discrepancies in the results for SSS and SST, particularly the higher variability of the SST data, require further clarification. Are these methodological differences or inherent in the data?

Response 3: This has been further explained in Section 3.1 (Lines 420-436).

Comment 4: The conclusions are too general and do not fully address the impact of methodological limitations on the results. The need for larger data sets and improved models should be emphasized. I suggest a more critical approach to the results, including specific recommendations for methodological improvements to improve prediction accuracy.

Response 4: I have included a paragraph in the conclusion discussing potential future work, highlighting how it can build upon the current study and suggesting ways to enhance prediction accuracy (Lines 585-602).

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Overall, the revision has been well done.

The geographical coordinates of the study area are not at the central position, and the four corner values of the rectangle should be added to the figure.

Comments on the Quality of English Language

Corrections need.

Author Response

Comment: The geographical coordinates of the study area are not at the central position, and the four corner values of the rectangle should be added to the figure.

Response: Added the geographic coordinates to figures 3,4,5.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed most of my comments; however, they appear to have misunderstood one of my key points. My concern was not about the applicability of the model to other parameters but rather its generalizability and effectiveness across different geographic regions. While the study on the "RF inversion of SST and SSS" provides valuable insights, its focus on a small-scale case study limits the broader implications of the findings. Similarly, regarding the "monitoring of Maltese Islands SST and SSS," the discussion on temporal variations and the underlying driving mechanisms remains somewhat superficial. Nonetheless, the study demonstrates a simple yet effective approach in term of "RF inversion of SST and SSS in the Maltese Islands."

Author Response

comment: The authors have addressed most of my comments; however, they appear to have misunderstood one of my key points. My concern was not about the applicability of the model to other parameters but rather its generalizability and effectiveness across different geographic regions. While the study on the "RF inversion of SST and SSS" provides valuable insights, its focus on a small-scale case study limits the broader implications of the findings. Similarly, regarding the "monitoring of Maltese Islands SST and SSS," the discussion on temporal variations and the underlying driving mechanisms remains somewhat superficial. Nonetheless, the study demonstrates a simple yet effective approach in term of "RF inversion of SST and SSS in the Maltese Islands."

Response: Addressed this comment in section 4 between lines 584 to 599.

Back to TopTop