Application of a PLS-Augmented ANN Model for Retrieving Chlorophyll-a from Hyperspectral Data in Case 2 Waters of the Western Basin of Lake Erie

Round 1

Reviewer 1 Report

 

Review of: Application of PLS-Augmented ANN Model for Retrieving Chlorophyll-a from Hyperspectral Data in Case 2 Waters of the Western Basin of Lake Erie.

 

The authors were interested in developing an improved model for estimating algal chlorophyll concentrations in the waters of western Lake Erie by taking advantage of the wealth of spectral information contained in hyperspectral data.  In this study they combined an artificial neural network with output from partial least squares regression to estimate chlorophyll concentrations at sampled locations in western Lake Erie.  They compared their method’s results to those obtained with a band-ratio algorithm based on reflectances in the blue and green spectral regions, a band ratio algorithm based on reflectances in the red and near-infrared (NIR) spectral regions, and a PLS-only approach. Their proposed method gave the most precise results of the four methods with an RMSE of 1.22 µg/l.

            I reviewed this manuscript from the standpoint of a limnologist interested in the use of remote sensing to estimate chlorophyll concentrations in several lakes over a broad geographic area or in the same lake over long time periods.  I was impressed with their RMSE of 1.2 µg/l.  This is much more precise than most methods I have found in the literature. I was, however, disappointed that the discussion did not discuss the application of their methodology to the real-world situation where reflectance values were obtained from satellite-measured reflectance spectra rather than from a spectroradiometer connected to a pole extended from the side of a research vessel at each sampling site. Also, I am interested in knowing if this methodology requires recalibration when used in the same lake over time or in several different lakes on the same day.

 

Small points

 

Line 14

 

            Most journals do not allow for citations in an abstract.

 

Lines 159-160

 

            This is not a complete sentence.

 

Figure 8.

 

            I would have each of the four figures identified with the algorithm names used in Table 1.  I would also identify each figure with A, B, C, or D.

Author Response

Reviewer 1

Review of: Application of PLS-Augmented ANN Model for Retrieving Chlorophyll-a from Hyperspectral Data in Case 2 Waters of the Western Basin of Lake Erie.

 The authors were interested in developing an improved model for estimating algal chlorophyll concentrations in the waters of western Lake Erie by taking advantage of the wealth of spectral information contained in hyperspectral data.  In this study they combined an artificial neural network with output from partial least squares regression to estimate chlorophyll concentrations at sampled locations in western Lake Erie.  They compared their method’s results to those obtained with a band-ratio algorithm based on reflectance’s in the blue and green spectral regions, a band ratio algorithm based on reflectance in the red and near-infrared (NIR) spectral regions, and a PLS-only approach. Their proposed method gave the most precise results of the four methods with an RMSE of 1.22 µg/l.

I reviewed this manuscript from the standpoint of a limnologist interested in the use of remote sensing to estimate chlorophyll concentrations in several lakes over a broad geographic area or in the same lake over long time periods.  I was impressed with their RMSE of 1.2 µg/l.  This is much more precise than most methods I have found in the literature. I was, however, disappointed that the discussion did not discuss the application of their methodology to the real-world situation where reflectance values were obtained from satellite-measured reflectance spectra rather than from a spectroradiometer connected to a pole extended from the side of a research vessel at each sampling site. Also, I am interested in knowing if this methodology requires recalibration when used in the same lake over time or in several different lakes on the same day.

Thank you

Response to reviewer 1

Thank you very much for the positive feedback. Yes, the results from our approach has shown that the method is robust.

This is simply a test of various approaches. We have added language to the text to clarify that the goal of the study was not to develop a universally applicable, satellite-based PLS-ANN algorithm but to simply test the PLS-ANN approach in comparison to a few other models for estimating chl-a concentration in low-moderate chl-a concentrations. Testing the approach on an in situ dataset helps us gain a better understanding of the inner workings of the model prior to application to satellite data, which are subject to various sources of uncertainties absent in in situ data (e.g., due to inaccurate atmospheric correction). When more hyperspectral sensors offering routine global coverage such as PACE and SBG are launched, we will collect larger datasets and develop algorithms based on this approach that are broadly applicable to satellite data from multiple water bodies covering a wide range of biophysical conditions.

This paper uses data from optically diverse waters within the WBLE (near shore to offshore waters), and therefore the models may not require to be recalibrated over time. However, some degree of recalibration may be needed when applying the method to various water bodies, especially if they encompass a very wide range of biophysical conditions.  

Small points

Line 14

            Most journals do not allow for citations in an abstract.

Response to reviewer 1

Thank you for the comment. We have removed the citation from the abstract.

 Lines 159-160

             This is not a complete sentence.

Response to reviewer 1

Thank you for catching this. We’ve fixed this sentence as follows:

During late spring and early summer, river discharges are high, and significant amounts of terrestrial matter, including nutrients, are transported into the bay 

 Figure 8.

             I would have each of the four figures identified with the algorithm names used in Table 1.  I would also identify each figure with A, B, C, or D.

Response to reviewer 1

Thank you for your pointing out this. We have added the letters as per your suggestion.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

SEE ATTACHED FILE

 In the present paper, the authors used the partial least squares (PLS) and artificial neural networks (ANNs) for predicting chlorophyll-a (chl-a) concentration in the western basin of Lake Erie (WBLE).The proposed modelling strategy was done using hyperspectral. The authors use the output of the PLS (i.e., the PCA) as input to the ANN model and in total the comparison was done between four models: the PLS, the PLS-ANN, the Blue-green Model and the NIR-red Model. Obtained results revealed that the PLS-ANN was more accurate and exhibiting the best performances based on the R², RMSE and MAE reported in Table1 line 503.

 It is impossible to act positively to the present work for several raisons. Originality, overall contribution very limited, paper structure and overall results and discussion section.

 General Comments.

 1.      What the authors propose is well known, broadly discussed in the literature, and proposed since more than two decades. The contribution is very limited.

2.      The paper is badly structured which has further complicated the understanding of the manuscript. With poor literature review and missing research gap, it is hard to be convinced with what is proposed.

3.      Novelty and overall contribution and not justified by the authors.

4.      The combination of the PLS with the ANN is not justified.

5.      Results and discussion is incomplete and insufficient.

 Major’s comments

 1.      Description of the dataset and the overall presentation and organization of the results is unclear and incomplete.

2.      The readers cannot understand if the presented results are for training or for validation. For any modeling study, it is well known and mandatory to clearly present the results separately for training and for testing, and the testing phase is critical for better models evaluation and comparison. This is not done in the present study, which has significantly reduced the value and the contribution of the present travail.

3.      Section discussion is missing.

4.      Some results are doubtful. In table 1, the MAE cannot be higher than the RMSE. The results should be rechecked by the authors

 

 

 

 

Comments for author File: Comments.pdf

Author Response

Reviewer 2

In the present paper, the authors used the partial least squares (PLS) and artificial neural networks (ANNs) for predicting chlorophyll-a (chl-a) concentration in the western basin of Lake Erie (WBLE).The proposed modelling strategy was done using hyperspectral. The authors use the output of the PLS (i.e., the PCA) as input to the ANN model and in total the comparison was done between four models: the PLS, the PLS-ANN, the Blue-green Model and the NIR-red Model. Obtained results revealed that the PLS-ANN was more accurate and exhibiting the best performances based on the R², RMSE and MAE reported in Table1 line 503.

It is impossible to act positively to the present work for several raisons. Originality, overall contribution very limited, paper structure and overall results and discussion section.

Response to reviewer 2

Thank you for the comment. Currently, retrievals of water quality parameters (WQP; e.g., Chlorophyll-a) is primarily done using simple linear based band ratios. The contribution of this study is to exhibit the robustness of coupling linear and non-linear based models to capture the dynamics of WQPs in optically complex (turbid) Case 2 type waters. Our results help to elucidate the robustness of using combined linear (PLS) and non-linear machine learning technique (ANN) for retrieving WQPs from hyperspectral data, which, we think, is an important contribution, considering the upcoming launch of spaceborne hyperspectral sensors such as the PACE and SBG missions, which will provide regular global coverage of hyperspectral data.

We have clarified the text (the last two paragraphs under Introduction and under Conclusion) to provide further context to this study and describe its significance, which is to test the suitability of the PLS-ANN approach in low-moderate chl-a concentrations, which represent an area of transition from mesotrophic to eutrophic waters, where algorithms developed for oligotrophic waters or eutrophic waters typically struggle to yield accurate estimates.

We have also modified the structure to follow the sequence of Introduction – Data & Methods – Results & Discussion – Conclusion, which is a standard format.

 General Comments.

1. What the authors propose is well known, broadly discussed in the literature, and proposed since more than two decades. The contribution is very limited.

Response to reviewer 2

Thank you for the comment. Although applications of PLS and ANN have been independently discussed in various studies, in our opinion, the combined potential of PLS and ANN to account for both linear and non-linear relationships between dependent and independent variables have not been sufficiently exploited in the context of aquatic remote sensing. Indeed, Song et al. (2014) published results from applying the PLS-ANN approach to aquatic remote sensing data. However, that was for a very wide range of chl-a concentrations. In our experience, and as can also been seen from a few of the figures in Song et al. (2014), published performance metrics for algorithms parameterized using a very wide range of parameter concentrations can be potentially misleading in the sense that the algorithm may not yield similar accuracies when applied to a dataset with a narrower range of conditions. In this context, we wanted to test the suitability of the PLS-ANN approach for low-moderate chl-a concentrations, which we have noted is also a range where several algorithms struggle to yield accurate estimates. We believe that the successful demonstration of the suitability of the PLS-ANN approach for estimating low-moderate chl-a concentrations is a valuable contribution, which, we hope, will lead to further development of this approach towards an algorithm readily applicable to data from future hyperspectral spaceborne sensors such as the PACE and SBG missions. We have added more text in Introduction and Conclusion to describe the motivation of this study and its significance.

2. The paper is badly structured which has further complicated the understanding of the manuscript. With poor literature review and missing research gap, it is hard to be convinced with what is proposed.

We have modified the structure to follow the sequence of Introduction – Data & Methods – Results & Discussion – Conclusion, which is a standard format. We have added more content in the Introduction to add more citations and provide a clearer background and motivation for our study.

3. Novelty and overall contribution and not justified by the authors.

As stated in our response to the General Comment #1, we have added text (in Introduction and Conclusion) to describe the motivation of our study and the value of its contribution.

4. The combination of the PLS with the ANN is not justified.

Thanks for the comment. We have stated the benefit of combining the PLS and ANN methods to reduce multicollinearity and dimensionality in the data and handle nonlinearity in the data. We have emphasized this throughout the manuscript.

5. Results and discussion is incomplete and insufficient.

It is not clear what the reviewer specifically means by “incomplete and insufficient”. We have listed the error metrics used to analyze the results and have described each error metric. We have shown in the form of figures, tables, and text, the error metrics for each of the four approaches we have tested. We have also discussed the performance of the PLS-ANN in comparison to that of the other approaches in terms of the listed error metrics. Additional information on what the reviewer specifically sees as missing in our results and discussion would be helpful.

 Major’s comments

1. Description of the dataset and the overall presentation and organization of the results is unclear and incomplete.

We have modified the structure of the manuscript to follow the sequence of Introduction – Data & Methods – Results & Discussion – Conclusion, which is a standard format. We hope that this makes the manuscript flow in a more logical sequence.

2. The readers cannot understand if the presented results are for training or for validation. For any modeling study, it is well known and mandatory to clearly present the results separately for training and for testing, and the testing phase is critical for better models evaluation and comparison. This is not done in the present study, which has significantly reduced the value and the contribution of the present travail.

      Thanks for the comment. We have added a few sentences in the first paragraph under Results and Discussion to make it clear that we did not split the data into separate calibration and validation datasets because our dataset is rather small. The goal of this study is not to develop universally applicable PLS-ANN algorithms but to rather test the suitability of the approach for a small dataset with low-moderate chl-a concentrations. Therefore, we find it appropriate to apply each of the algorithms to the entire dataset and assess the retrievals by comparing them to in situ measurements.

3. Section discussion is missing.

Response to reviewer 2

Thanks for the comment. Please note that the results and discussion are combined into a single section. The Conclusion also contains some discussion.

4. Some results are doubtful. In table 1, the MAE cannot be higher than the RMSE. The results should be rechecked by the authors

Response to reviewer 2

Thank you.  Please note that as indicated on lines 321-322 and defined in equations 15, 16 and 17, the MAE we have calculated is based on log10-transformed values and are unitless measures, while the RMSE is based on untransformed values and are in physical units of chl-a concentration. Hence, they cannot be directly compared to each other. For reference, the  MAE for the PLS-ANN method based on untransformed values is 0.913 ug/l, which is lower than the RMSE. The approach we have followed with the presentation of the error matrix is consistent with other studies in ocean color remote sensing  (e.g., Pahlevan et al., 2020).

 

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors,

The article you presented tries to prove that PLS-ANN model is  a good approach for calculating chlorophyll-a from hyperspectral data. You do that by testing different approaches (4) with in situ reflectance measurements on Lake Erie. 

You say that Lake Erie is an optically complex water body but for the readers its not clear why. You do not explain it anywhere. It is also not clear what was the criteria for choosing in situ measurement locations? Do they represent the complexity of Lake Erie? Methods and material section has to be improved with information why Lake Erie was chosen and what were the criteria for choosing reference measurement locations.

In the introduction part you state the Song et al have already tested similar PLS-ANN approach for extracting chlorophyll-a (on several inland water bodies!). What is the contribution of your work then? Also, in the results section for PLS-ANN you do not compare your results with Song et al to see if they are similar.

Article is missing a discussion part. What would be the next steps? Are you planning to use satellite imagery and test PLS-ANN approach on that data? In the introduction part you have said that there are several hyperspectral mission where to obtain data in the near future (PRISMA and EnMAP).

 

Author Response

Reviewer 3

The article you presented tries to prove that PLS-ANN model is a good approach for calculating chlorophyll-a from hyperspectral data. You do that by testing different approaches (4) with in situ reflectance measurements on Lake Erie. 

Comment #1: You say that Lake Erie is an optically complex water body but for the readers its not clear why. You do not explain it anywhere. It is also not clear what was the criteria for choosing in situ measurement locations? Do they represent the complexity of Lake Erie? Methods and material section has to be improved with information why Lake Erie was chosen and what were the criteria for choosing reference measurement locations.

Response to reviewer 3

Thank you for the comment. In line 46, we briefly describe what we mean by ‘optically complex waters’, which is that the optical signatures of such waters are a function of multiple non-covarying water quality parameters (i.e., it is not dominated by just a single parameter). Lake Erie is known as an optically complex water body. We have added the following reference (reference [9] in line 46) to support the claim:

[9] D. M. O'Donnell, S. W. Effler, C. M. Strait, and G. A. Leshkevich, “Optical characterizations and pursuit of optical closure for the western basin of Lake Erie through in situ measurements,“ J. Great Lakes Res., vol. 36, no. 4, Dec. 2010, pp. 736-746, doi: 10.1016/j.jglr.2010.08.009.

Comment #2: In the introduction part you state the Song et al have already tested similar PLS-ANN approach for extracting chlorophyll-a (on several inland water bodies!). What is the contribution of your work then? Also, in the results section for PLS-ANN you do not compare your results with Song et al to see if they are similar.

Response to reviewer 3

Thank you for the comment. We have added text in the last paragraph under Introduction and the first paragraph under Conclusion to clarify the distinction between our study and that of Song et al. (2014). Song et al. (2014) published results from applying the PLS-ANN approach to aquatic remote sensing data. However, that was for a very wide range of chl-a concentrations. In our experience, and as can also been seen from a few of the figures in Song et al. (2014), published performance metrics for algorithms parameterized using a very wide range of parameter concentrations can be potentially misleading in the sense that the algorithm may not yield similar accuracies when applied to a dataset with a narrower range of conditions. In this context, we wanted to test the suitability of the PLS-ANN approach for low-moderate chl-a concentrations, which we have noted is also a range where several algorithms struggle to yield accurate estimates. We believe that the successful demonstration of the suitability of the PLS-ANN approach for estimating low-moderate chl-a concentrations is a valuable contribution, which, we hope, will lead to further development of this approach towards an algorithm readily applicable to data from future hyperspectral spaceborne sensors such as the PACE and SBG missions.

We have also added a sentence in the first paragraph under Conclusion to state that our results are consistent with those from Song et al. (2014) in that the PLS-ANN model performed better than the band-ratio algorithms.

Comment #3: Article is missing a discussion part. What would be the next steps? Are you planning to use satellite imagery and test PLS-ANN approach on that data? In the introduction part you have said that there are several hyperspectral missions where to obtain data in the near future (PRISMA and EnMAP).

Response to reviewer 3

The discussion is combined with the results in the Results and Discussion section. We have also expanded the Conclusion to add more discussion. In Conclusion, we have stated that the next step would be to collect a much larger dataset encompassing a wide range of biophysical conditions in order to develop PLS-ANN-based algorithms for broader application to hyperspectral satellite from data multiple water bodies without the need for site-specific calibration. A key consideration in developing such a broadly applicable algorithm would be consistency in the spectral channels associated with the most significant PLS factors.  

 

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The actual form of the paper is more than the first one, and the authors have worked seriously for improving the manuscript. The paper can now be accepted, no further revision is necessary.

Reviewer 3 Report

The comments are sufficient for accepting the paper.