Adaptation of a Neuro-Variational Algorithm from SeaWiFS to MODIS-Aqua Sensor for the Determination of Atmospheric and Oceanic Variables

Round 1

Reviewer 1 Report

 

 

In their recent study, Correa et al. explore the adaptation of a two-step machine learning algorithm to address the challenging task of atmospheric correction in MODIS-Aqua images. The paper aims to tackle a significant obstacle in this field, namely the influence of absorbing aerosols on the retrieval of accurate ocean color properties. As a reviewer, I agreed to assess this paper due to its relevance to the fundamental question of aerosol impacts on ocean color measurements. While my expertise lies outside the realm of machine learning techniques, which are not the primary focus of this study, I have provided some suggestions for improvement based on my understanding.

Although the study by Correa et al. does not present groundbreaking advancements in the characterization and consideration of aerosols for atmospheric correction, it does propose a viable treatment for MODIS data. It is important to note that keeping up with the latest advancements in this rapidly evolving field can be challenging during the process of writing a scientific paper. Nevertheless, I recommend that the authors familiarize themselves with the recent publications by Song et al. 2022 (https://doi.org/10.1109/TGRS.2022.3151219) and 2023 (https://doi.org/10.1016/j.rse.2023.113552), which provide valuable insights to enhance their study.

One aspect that warrants further clarification is the focus solely on NIR radiances in the context of absorbing aerosols and considering their significance across the entire spectrum. Addressing this point could strengthen the authors' analysis and contribute to a more comprehensive understanding of the subject matter.

Consequently, I advise the authors to revise the manuscript and submit an improved version to gain in readability and to enhance the result analysis that would enhanced readers interests. The materials is present to propose a publishable study. Furthermore, the addition of a dedicated discussion section would greatly enhance the flow and structure of the paper.

Detailed comments can be found in the attached .pdf

Comments for author File: Comments.pdf

Author Response

Reviewer #1

 

Review MDPI Remote Sensing Correa et al 2023 : Adaptation of a Neuro-variational algorithm from SeaWiFS to MODIS-Aqua sensor for the determination of atmospheric oceanic variables

 

General comments

In their recent study, Correa et al. explore the adaptation of a two-step machine learning algorithm to address the challenging task of atmospheric correction in MODIS-Aqua images. The paper aims to tackle a significant obstacle in this field, namely the influence of absorbing aerosols on the retrieval of accurate ocean color properties. As a reviewer, I agreed to assess this paper due to its relevance to the fundamental question of aerosol impacts on ocean color measurements. While my expertise lies outside the realm of machine learning techniques, which are not the primary focus of this study, I have provided some suggestions for improvement based on my understanding.

Although the study by Correa et al. does not present groundbreaking advancements in the characterization and consideration of aerosols for atmospheric correction, it does propose a viable treatment for MODIS data. It is important to note that keeping up with the latest advancements in this rapidly evolving field can be challenging during the process of writing a scientific paper. Nevertheless, I recommend that the authors familiarize them-selves with the recent publications by Song et al. 2022 (https://doi.org/10.1109/TGRS.2022.3151219) and 2023 (https://doi.org/10.1016/j.rse.2023.113552), which provide valuable insights to enhance their study.

One aspect that warrants further clarification is the focus solely on NIR radiances in the context of absorbing aerosols and considering their significance across the entire spectrum. Addressing this point could strengthen the authors’ analysis and contribute to a more comprehensive understanding of the subject matter.

Consequently, I advise the authors to revise the manuscript and submit an improved version to gain in readability and to enhance the result analysis that would enhanced readers interests. The materials is present to propose a publishable study. Furthermore, the addition of a dedicated discussion section would greatly enhance the flow and structure of the paper.

We would like to thank reviewer #1 for his pertinent and constructive comments. We also thank him for the valuable references provided. We have taken the suggestions for improvement into account and hope to have responded satisfactorily to all the clarifications and comments made. In line with these, we have rewritten the introduction and written a new discussion section.

 

Introduction

Upon reviewing the manuscript, I found that the introduction lacked readability and could be improved by focusing on the primary objective of the study. To enhance clarity, I suggest emphasizing the main purpose of the research, which revolves around the impact of absorbing aerosols on atmospheric correction in ocean color remote sensing. Providing a comprehensive overview of the current state-of-the-art in this specific domain would be beneficial. For instance, instead of starting the introduction with the importance of ocean color in studying small-scale spatiotemporal phytoplankton dynamics, which, although valid, is not directly related to the study’s domain, it would be more effective to focus on the relevance of absorbing aerosols in atmospheric correction. This would provide a clear and concise introduction to the core topic of the paper.

Additionally, it would be advantageous to include a description of the current state-of-the-art regarding the utilization of machine learning techniques in retrieving aerosol properties and improving atmospheric correction for oceanic outcomes. This would provide context and background information for readers, allowing them to understand the significance and novelty of the study.

Regarding the list of references provided on line 35, it would be helpful to provide some contextual information or brief explanations regarding how these references are related to the topic at hand. This would aid readers in understanding the relevance and importance of these cited works in the context of the present study.

In the section from line 39 to 65, I encountered difficulties in following the purpose of the discussion. The description initially focused on the impact of absorbing aerosols on remote sensing measurements, then shifted to discussing the regional significance of these properties, and subsequently returned to addressing the approach taken to overcome issues in atmospheric correction. To improve the coherence and flow of this section, I recommend rewriting it.

Conventional atmospheric correction procedures are commonly based on the aerosol contribution in the NIR and SWIR bands, with the water signal being considered as a “dark pixel” (as the authors also mentioned in the Materials and Methods section). However, it is important to note that these methods may not work effectively in coastal regions despite a portion of the study area not being exclusively coastal.

Furthermore, the impact of absorbing aerosols in other bands, such as the blue band, which holds significant importance (especially since you utilize all 8 bands of SeaWifs and 7 bands of MODIS), should be addressed.

Lastly, it is crucial to highlight the main limitations of current atmospheric correction procedures, which are linked to the under-representation of the vertical distribution and type diversity of absorbing aerosols. Introducing these limitations would provide valuable insights into the challenges faced in the field and further underscore the significance of the proposed research. By addressing these points, the overall readability, coherence, and contextual understanding of the paper will be greatly improved.

Following the many constructive suggestions made by Reviewer #1, we have reorganized the introduction and added the elements needed to make this section more coherent and fluid. We now hope to emphasize the importance of absorbing aerosols in the atmospheric correction of water color signals. Building on this, we have also written a state-of-the-art review of the correction principle applied to SeaWiFS and MODIS sensors, starting from the basic algorithm of Gordon and Wang (1994) and describing initiatives to use machine learning methods most relevant to our study region. In this respect, we would like to thank Reviewer #1 for having drawn our attention to the work of Song and co-authors, which turned out to be highly relevant and had slipped under the radar of our literature watch.

In this way, we hope to have better defined the contextual framework, improved the readability of this introduction and its coherence.

 

 

Materials and Methods

L76:

I noticed a potential typo in the text. It appears that you meant to refer to "MODIS-Aqua L1B data" instead of just "MODIS-Aqua L2 data" On line 83, you mentioned AC to L2 retrieve products.

We're talking about MODIS level 2 data here. This corresponds to the standard data that we then use to evaluate the performance of the SOM-NV algorithm. We've added this sentence "for comparison with the processing proposed in the present study" (L113-114) to clarify the reason for downloading these level 2 data. In the section 2.2 we've added the sentence "from which the SOM-NV processing starts" (L166-167) to emphasize the difference between STD and SOM-NV processing. In the SOM-NV approach, we download the Level 1 data and apply the standard atmospheric correction to the left-hand terms of equation 4.

 

 

Results

I found this section to be interesting; however, the results lack in-depth analysis. To increase the overall interest and comprehensibility of this section, I suggest incorporating metrics that can effectively evaluate the performance of the algorithm employed.

Specifically, in the section discussing AOT, the authors only employ the Pearson coefficient as a measure of correlation. While this provides some insight, it would be beneficial to include additional evaluation metrics that can assess the performance and accuracy of the algorithm in a more comprehensive manner. This would strengthen the analysis and provide a more robust interpretation of the results.

To better highlight SOM-NV's performance in determining the AOT, we first added the number of points used for both stations and processing. We also calculated the regression between the AOTs of the two products STD and SOM-NV and those of AERONET, and show the lines obtained in the new Figure 1. We discuss these statistical relationships in subsection 3.1.

 

Furthermore, the analysis of seasonality is currently presented solely through visual means. It would be valuable to supplement this with quantitative analysis to support the visual observations and provide a deeper understanding of the seasonality patterns.

To reinforce visual appreciation and better understand the method's performance by season, we have added a table of correlations (and significance) between the AERONET AOT and the STD and SOM-NV AOTs for each season. These correlations are presented in section 3.1, lines 315 to 318 (plus Table 2).

 

In the data coverage section, the analysis is primarily visual, lacking the inclusion of relevant evaluation metrics. While it is evident that the data coverage has improved, it would be valuable to assess the statistical correctness and reliability of the data. I recommend incorporating an evaluation of the relevance and statistical soundness of the extended data coverage to strengthen the conclusions drawn from this section. This additional analysis would contribute to the overall persuasiveness of the results (this point is of particular importance knowing about the results of the validation section afterwards).

Coverage is not an intrinsic parameter of the SOM-NV method. Pixels discarded by SOM-NV result from the failure of the minimization step only. STD coverage has a certain level of arbitrariness, as we could recover pixels by relaxing some of the masks taken into account. Nevertheless, Figure 2 illustrates that the sub-areas where pixels are considered by SOM-NV (and not by STD) show coherent Chla concentrations and spatial structures that remain independent of the shape of the masked areas and in agreement with the known dynamics of the study area.

We have added Figure 3 showing the Chla distribution for this January 4, 2003 image. We show the distribution of all the SOM-NV pixels in this image, the pixels in common with the standard image (by applying the STD mask) and finally the pixels treated by SOM-NV but not by STD. The distribution for pixels processed by SOM-NV but not by STD does not differ from that of all pixels or from that of pixels processed by both. The chi-square test shows that these different distributions match. We hope that this statistical evaluation, carried out on a day when dust is clearly present, will reinforce the reliability of the data generated by SOM-NV.

 

Overall, it appears that the full potential of the results has not been fully explored, which may impact the findings. By leveraging a wider range of analytical techniques and evaluation metrics, the authors can enhance the overall persuasiveness and significance of the results obtained.

In agreement with Reviewer #1, we have added a number of evaluation metrics, which we hope will meet the reviewer's expectations.

 

L256

The size distribution of aerosol is one of the most important properties in atmospheric and also one the main limitation in retrieving relevant Rrs. This should be presented in the section Material and Method.

In line with this comment, the lines 256-259 concerning the size distribution considered by SOM-NV have been moved to section 2.2.1 (new lines 195-200).

 

L268-277

This section is not a result but a discussion and should be moved to the corresponding section.

This section has been moved to a new discussion section accordingly.

 

L279-284

This section should be part of the Materials and methods.

The calculation of the coverage is now defined in subsection 2.2.3 in the Material and methods section (L248-256).

 

L285

“the SOM-NV algorithm better takes into account absorbing desert dust aerosols’ This is not the result of the previous section. The previous section evaluate the algorithm in retrieving AOT on a single wavelength. I’m confident that your work might result in such an improvement but you have to show result that can confirm this.

A shortcut has indeed been taken in the wording of this sentence. Diouf et al (2013) showed that the SOM-A-S component of the SOM-NV algorithm was capable of determining different types of aerosols and dust in particular (e.g. their Fig. 5d). We rely on this result of the method applied to our region of interest. There is no physical reason why using the 531 nm wavelength (MODIS) instead of 510 nm (SeaWiFS) should alter the detection of desert dust. The improvement in AOT estimation results from the better determination of aerosol types. This implies that SOM-NV enables us to better characterize the optical effects of these desert dust particles, which represent a large proportion of the absorbing aerosols in our study region. The sentence has been reworded accordingly (L320-324).

 

L295

‘has improved significantly’

It seems that the authors performed a visual analysis. However, it is important to note that the use of statistical vocabulary should be avoided in this context.

We agree with the reviewer, the sentence has been reworded to avoid using statistical vocabulary here. ‘has improved significantly the data coverage’ has been reworded into ‘has increased the data coverage’ (L350).

 

L301-307

You are not presenting results here and this is part of the discussion.

This paragraph has been moved to the new discussion section.

 

L309

What kind of sea surface temperature do you use? What is the source? Reanalysis, analysis?

In addition, the Figure A1 is not a SST distribution.

Figure A1 is indeed a map of averaged MODIS SST over the years 2003-2015. It was computed by co-author H. Demarcq. We have changed ‘distribution’ to ‘map’ and added the missing information on the origin of the product.

 

L309-318

I concur with your explanation; however, it appears to be more of an assumption rather than a conclusive result. To substantiate your argument, I suggest incorporating an atmospheric analysis to validate your point. As it stands, your current presentation lacks tangible findings and primarily offers potential explanations for the performance of the SOM-NV algorithm.

Although coverage is higher off Cape Blanc (~20°N) than off southern Senegal, SOM-NV recovers more pixels in the south than in the north (Figure 4). In an attempt to strengthen the explanation for this pattern, we show an average 2003-2018 map of the number of SST MODIS observations. Fewer observations in the south indicate a complex atmosphere that makes it difficult to exploit the marine signal (Figure A1b). According to Pfah and Sodemann (2014), relative humidity is between 70 and 90% in the region. This humidity shifts the particle size spectrum towards larger particles and changes the refractive index. As a result of north-eastward dominant winds, desert dust is also more prevalent in the southern part of the plume. All these factors combine to make the southern part effectively more complex to correct, and indicate that the SOM-NV method seems to perform better. A full statistical analysis of atmospheric parameters is beyond the scope of this paper. Nevertheless, accepting the lack of compelling argument, we have chosen to move this amended explanation to the new discussion section.

 

 

L333-337

Not a result, should be moved either to the introduction or discussion sections.

As this paragraph doesn't provide any additional information where it belongs, we've chosen to keep it slightly reworded as an introduction to the discussion section (L427-433).

 

 

L338-351

Once again, this section lacks the presentation of any specific results. Instead, references are used to explain reflectance variability. Additionally, the utilization of different wavelengths is mentioned, but no analysis of the obtained results is provided. It would be beneficial to explain the discrepancies between the STD and SOM-NV results, particularly with respect to the response in the blue band. One possible explanation for the disparity could be the unsuitability of the NIR/SWIR approach for retrieving Rrs in regions with high concentrations of absorbing aerosols, which may lead to overestimation of Rrs (as described in IOCCG 2010). This issue could be further connected to aerosol distribution and the vertical profile.

We now discuss the difference in mean reflectance at the five wavelengths as a function of chlorophyll a concentration for the two treatments in the new discussion section (L462-476).

 

Furthermore, the analysis of Chl-a presented by the authors is not sufficiently convincing. Once again, the authors rely on visual interpretation of maps without employing any quantitative metrics. The statement that Chl-a concentrations differ between STD and SOM-NV lacks validation to determine which one more accurately reflects reality.

We have introduced new metrics to quantify the performance of the two treatments in determining Chl-a concentration. These are discussed in the section "Validation of Chla estimated by SOM-NV".

 

 

L365-368

Can you add a more quantitative analysis

We show the seasonal mean maps with an identical color scale, as well as sections on the latitude bands 18°-18.5°N and 13°-13.5°N. The fact of having colors associated with higher concentrations for SOM-NV and seeing the curves superimposed is a quantitative way of showing the differences between the two treatments. We therefore fail to understand this comment.

 

L355-378

Most of this section is a discussion and does not present any results.

Most of this section has been moved to the new discussion section.

 

L396-403

I would recommend to move this section to the Methods sections Validation of Chl-a estimated by SOM-NV sections As you correctly pointed out, the sample size is too small to yield statistically significant results. Additionally, the data has not been validated.

In agreement with Reviewer #1, these lines have been moved to the end of section 2.1 (L132-141).

 

Therefore, I strongly recommend conducting an additional validation procedure. One approach could involve comparing your results with those obtained using other published methods or utilizing a radiative transfer model as a baseline. This comparative analysis would provide a more robust validation of your findings and enhance the credibility of your study.

In the literature, we find studies for case 1 waters where correlations between in situ observations and satellite estimates are better than those presented here (e.g. Carswell et al., 2017) and others associated with case 2 coastal waters where regressions are not necessarily better (e.g. Abbas et al., 2022). The problem with our study region is that two difficulties overlap, i.e. the very frequent presence of desert dust and a coastal upwelling system where concentrations are high (but not necessarily turbid, as it borders dry regions where runoffs are rare or low). The lack of observation of this southern region of the Canary Islands system is regularly highlighted in reviews of the system (Aristegui et al., 2009; Kudela et al., 2010 ; Trainer et al., 2010). So we don't know how best to convince ourselves of the reliability of Chla derived from SOM-NV processing.

 

Abbas, M. M., Melesse, A. M., Scinto, L. J., & Rehage, J. S. (2019). Satellite estimation of chlorophyll-a using moderate resolution imaging spectroradiometer (MODIS) sensor in shallow coastal water bodies: Validation and improvement. Water, 11(8), 1621.

Arístegui, J., Barton, E. D., Álvarez-Salgado, X. A., Santos, A. M. P., Figueiras, F. G., Kifani, S., ... & Demarcq, H. (2009). Sub-regional ecosystem variability in the Canary Current upwelling. Progress in Oceanography, 83(1-4), 33-48.

Kudela, R.M., Seeyave, S., W.P. Cochlan, 2010. The role of nutrients in regulation and promotion of harmful algal blooms in upwelling systems. Progress in Oceanography, 85, 122135.

Trainer, V.L., G.C. Pitcher, B. Reguera, T.J. Smayda, 2010. The distribution and impacts of harmful algal bloom species in eastern boundary upwelling systems. Progress in Oceanography, 85, 3352

 

 

Minor comments

L22

I would not recommend using ’synoptic’ when referring to spatial sampling properties. Instead, I suggest using ’large scale’ or other synonymous terms. ’Synoptic’ has a specific definition that is associated with both spatial and temporal scales.

We have changed the term 'synoptic' to 'large scale' (L20).

 

L25

NAXA → JAXA

Typing error corrected

 

L355

Characterize → characterise

This US English error has been corrected, as well as in other places.

Author Response File: Author Response.pdf

Reviewer 2 Report

Atmospheric correction is essential and difficult for satellite ocean color remote sensing especially with the presence of highly absorbing aerosols. This manuscript presented an updated algorithm to perform such correction in the northwest of Africa. Overall, this article is worth publication. But also it is obvious with weakness. There is a lack of solid evidence that the new algorithm does perform better than the standard approach. Below I have provided some specific questions with a hope to help polish this manuscript.

Specific questions:

Q1: Section 3.1 --- SOM-NV obviously overestimated AOT

Q2: Section 3.1 --- The correlation coefficient is not a good metric as it is sensitive to the data range and relatively larger values

Q3: Figure 1a and Figure 2a --- Why the numbers and ranges of the AOT(870) data from aeronet are different (red versus blue)?

Q4: The first paragraph in Section 3.3 is superfluous --- it should belong to the introdution

Q5: The authors concluded improved reflectance spectra based on the gradual variation of Rrs and resulting Chla. This is rather subjective as the Rrs-Chla relationship is expected to be more complex.

Q6: Following the above question, in Figure 4, Rrs from SOM-NV seems quite questionable. For example, when Chla=0.1 mg m^3, the Rrs spectrum would unlikely reach nearly 0.1 sr^-1 at 410 nm.  I understand the authors probably did not have in situ Rrs measurements for validations. But several innovative approaches have emerged in last several years, which can be used to assess the Rrs data quality. I would suggest the authors to give them a try, such as the quality assurance score system.

Author Response

Reviewer #2

 

Q1: Section 3.1 --- SOM-NV obviously overestimated AOT

We have added the regression lines between the STD, SOM-NV products and AERONET observations. SOM-NV does overestimate daily AOT, more so at the edge of the continent (Mbour Station) than at Cabo Verde. Nevertheless, the correlation between SOM-NV and AERONET is much more satisfactory than the STD product. It should also be noted that AERONET averages are daily averages and that intraday variability can be significant (example of variation for the randomly chosen date of March 04, 2014: https://aeronet.gsfc.nasa.gov/cgi-bin/data_display_aod_v3?site=GSFC&nachal=0&year=2014&month=3&day=4&aero_water=0&level=1&if_day=0&if_err=0&place_code=10&year_or_month=0). We have added a sentence in the manuscript to mention this. It should be noted that the monthly average of MODIS AOTs remains lower than the monthly AERONET product (Fig.1c).

 

Q2: Section 3.1 --- The correlation coefficient is not a good metric as it is sensitive to the data range and relatively larger values

We agree with Reviewer #2. The correlation coefficient here is indicative, and we have given more credence to the p-value.

 

Q3: Figure 1a and Figure 2a --- Why the numbers and ranges of the AOT(870) data from aeronet are different (red versus blue)?

In Figures 1a and 1b, we have considered SOM-NV estimates for a 10 km radius around the AERONET stations. The same treatment has been applied to STD. Given that AOTs are thresholded at 0.3 in the standard product, unlike the SOM-NV product, it is normal for the number of points to be lower for the STD product. This was also the case for SeaWiFS ocean color data (Fig. 6 in Diouf et al., 2013).

 

Q4: The first paragraph in Section 3.3 is superfluous --- it should belong to the introdution

We agree with Reviewer #2, the message associated with this paragraph has been moved to the introduction.

 

Q5: The authors concluded improved reflectance spectra based on the gradual variation of Rrs and resulting Chla. This is rather subjective as the Rrs-Chla relationship is expected to be more complex.

We do not fully understand this remark. For Case 1 waters, the relationship between Chla concentration and reflectance spectrum shown in Figure 4 is in agreement with the reference model published by Morel and Maritorena (2001) for the same concentration range (Fig. 8). This spectral response is also in agreement with the relationship found by Alvain et al. (2008 ; Fig. 1), the mean spectra were obtained from 28,800 SeaWiFS spectra coinciding with Chla concentration and water-leaving radiance spectra located near the GeP&CO (Geochemistry, Phytoplankton, and Color of the Ocean program) ship tracks.

 

Q6: Following the above question, in Figure 4, Rrs from SOM-NV seems quite questionable. For example, when Chla=0.1 mg m^3, the Rrs spectrum would unlikely reach nearly 0.1 sr^-1 at 410 nm.  I understand the authors probably did not have in situ Rrs measurements for validations. But several innovative approaches have emerged in last several years, which can be used to assess the Rrs data quality. I would suggest the authors to give them a try, such as the quality assurance score system.

In fact, in situ Rrs do not exist for this validation. Again, we fail to understand why the Rrs should not reach 0.1 sr^-1 at 410 nm, a value close to that defined theoretically by Morel and Maritorena (2001) and to that derived from SeaWiFS sensor data (Alvain et al., 2005; Fig. 1). We thank Reviewer #2 for the information on the availability of a quality assurance score system for Rrs data, but we do not share the reservations expressed about their quality.

Round 2

Reviewer 1 Report

Thank you to the authors for considering and incorporating most of the comments in this updated version. The revised introduction and discussion sections are now comprehensive and greatly assist the reader.

I have just noticed some discrepancies in the references of the updated manuscript that I received. I'm uncertain about the source of these errors, and there are also some missing references indicated by question marks ('?') in the manuscript. It seems there may have been an issue during the compilation process (that explains the minor revisions).

Overall, this version fulfills the requirements for publication.