Machine and Deep Learning Regression of Chlorophyll-a Concentrations in Lakes Using PRISMA Satellite Hyperspectral Imagery

Round 1

Reviewer 1 Report

The paper entitled “Machine and Deep Learning Regression of Chlorophyll-a Concentrations in Lakes Using PRISMA Satellite Hyperspectral Imagery” applies the machine learning and deep learning algorithms to extract Chlorophyll-a concentration in inland lakes via hyperspectral images. For now, the paper needs to be revised and is far from being published in the journal.

Specific comments:

1.    The research used the product of chlorophyll-a generated by SIMILE project derived from Sentinel 3 as reference data (Line 127-128), and mentioned that SIMILE has in-situ measurements (Line 115-116). However, is there any more direct result which prove that SIMILE’s Chlorophyll-a product can be treated as true values.

2.    In the introduction section, the authors mentioned that the chlorophyll-a concentration map of Sentinel 3 would be reconstructed (Line 140-141), but in Section 2.4, it is not mentioned how it was reconstructed, and whether the reconstructed results are more reliable.

3.    Why only half of the lake Maggiore data in the study area is data-covered (Figure 1)?

4.    The QQplot shown in Figure 4 still reflects that the overall number of samples is small. In neither machine learning nor deep learning, when the sample size is too small and only over-tunning hyperparameters, how to ensure the models’ accuracy and not over-fitting?

5.    The image processing for PRISMA is mentioned in this paper, but in the training of the model in Section 2.4, there is no detailed description of the input data. Whether the reflectance of PRISMA bands corresponding to Sentinel3 is used as input or the reflectance of all bands of PRISMA is used as input?

6.    In the part of RF model performance, only one set of control experiments between RF10 and RF12 were conducted to do the effect of image resolution on the model’s accuracy, but it is difficult to ensure that the consistency of other parameters in this set (Line 345-346). Under such conditions, the final conclusion that RF12 has a higher error seems to be not convincing (Line 347-348). Also, how to determine the hyperparameters? Random search cross-validation?

7.    How to explain that the application of the PCA strategy works poorly in the RF model, but works best in the SVR?

8.    In Figure 5, it is evident that the best machine learning model’s predictions undervalued at the relative high chlorophyll-a concentration region in lakes generally.

9.    The paper does not reflect the advantages of hyperspectral images in the retrieval of typical water parameters.

Author Response

Dear Reviewer,

I would like to express my sincere gratitude for dedicating your valuable time to reviewing this manuscript. Please find the detailed responses in the attached PDF file.

Best regards,

Juan Francisco Amieva.

Author Response File: Author Response.pdf

Reviewer 2 Report

The article entitled “Machine and Deep Learning Regression of Chlorophyll-a Concentrations in Lakes Using PRISMA Satellite Hyperspectral Imagery” provides an interesting approach to modern methods in water quality detection. I am convinced in using the proposed workflow for researching the seasonal variability of chlorophyll a. My doubts are about the low spatial resolution (300 m) of input materials (even though it is resampled and characterized in high - 2 days time resolution). After reading this article I would say that the presented research is more technical than scientific. But considering the wide possibility of using proposed methods, in my opinion, after minor revision this article should be allowed for publication.

Referring to the article's language for me everything is understandable, with no English mistakes, although I am not feeling qualified to assess it.

Below, I am referring to a particular section of the paper advising on what should be improved.

Abstract

The abstract is well written, so I haven’t got any comments on this part. Both the abstract and the entire structure of the article are correct - consistent and followed by rules assigned to scientific articles.

Introduction

Line 20-22: In the first sentence it should be mentioned that the influence of natural factors on eutrophication is negligible and then it should be highlighted that anthropopressure significantly intensifies eutrophication

Line 22-24: In this line as well as in the next part of the text authors refer just to remote research on lakes. In the second part of the text, some of the models are presented also for other types of waters. So it should be mentioned from the beginning that remote detection is applied also in research on other inland and marine water. Especially since there are many research, which You can refer eg. DOI: 10.1016/j.ecolind.2023.110103.

Line 35-54: In this paragraph, there is a lack of explanation as to why is chlorophyll very well detected by sensors and how the reflectance process is working.

Line 98 – 103: The authors should be appreciated for noticing that 300 m spatial resolution is very low but make it possible to have a 2-day time resolution and one image for a wide area.

Line 109 – 122: In this paragraph or next section there is a lack of explanations as to why these 3 lakes were taken into account (not others)? Actually, one sentence that they are representative for the project area should be enough.

Line 140 – 148: The resampling process helps to unify data for further processing but data quality isn’t better. Was it considered in Your research that data taken from 300 m resolution images is very generalized and gives overall information (we miss areas with high point chlorophyll concentration)?

Data and methods

Lines 157 – 170: It should be explained if every image (taken in 2- and 29-day intervals) was used or if some of them were rejected (e.g. because of high cloudiness). 

Line 218 and 225: I suppose that this pixel algebra is classified as some indicators (e.g. Vegetation Index). It should be checked and If these mathematical operation was named by some index it should be mentioned.

Results and Discussion:

The concise way of discussion of the research method and results deserves praise. The shown workflow makes it possible for readers to understand what data was used and why the given results are fine or not. Then this research could be repeated on other data. As I mentioned before  I think that even a proper model (shown in the paper) with low-resolution input data won’t be good enough. The resampling process doesn’t improve spatial resolution (300 m). High generalized data were easier to test and gave more consistent results because they were less detailed, which authors noticed in lines 402-405, but it could be more highlighted or better explained. I left it just to be considered by authors.

For sure presented workflow could be used for more accurate data and that is why I think that this chapter is crucial for further research.

Conclusion:

This section is written very clearly and even if it is short it contains everything necessary for summary and conclusion, so I haven’t got any comments.

Author Response

Dear Reviewer,

I would like to express my sincere gratitude for dedicating your valuable time to reviewing this manuscript. Please find the detailed responses in the attached PDF file.

Best regards,

Juan Francisco Amieva.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors apply various machine learning techniques to estimate chlorophyll (chla) concentrations using hyperspectral data as input. The study is well designed and implemented, and the writing is good overall (see below and pdf for comments).  The main weakness is the lack of “raw” ground reference data---the reference dataset used for training, validation and testing (SIMILE) is itself based on remote sensing, so differences between the hyperspectral product and SIMILE could be errors in SIMILE.  What is the error in SIMILE compared with ground reference data, and how does that compare with the errors (or differences) between SIMILE and hyperspectral-based model?

Specific comments:

L30.  SGD here seems a bit of a stretch…isn’t there an SGD on ecosystems?

The literature review to L91 is very comprehensive.

Style note:  “-“ are not commonly used; instead, use a “,”

“largely” is often used in the manuscript, but isn’t really the right word.  “Often” or “frequently” are more accurate and grammatically correct.

See additional comments in the pdf, where “awk” = awkward wording; “wc” = word choice error, “?” = confusing or unclear.

Comments for author File: Comments.pdf

The writing is good overall (but see other comments in the evaluation and pdf)

Author Response

Dear Reviewer,

I would like to express my sincere gratitude for dedicating your valuable time to review this manuscript. Please find the detailed responses in the attached PDF file.

Best regards,

Juan Francisco Amieva.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

authors addressed all the comments from me and I have no more comments.