A Predictive Model of Chlorophyll a in Western Lake Erie Based on Artificial Neural Network

Round 1

Reviewer 1 Report

Review: A Predictive Model of Chlorophyll a in Western Lake Erie Based on Artificial Neural Network

This is overall a well written, concise and solid scientific study and I believe it is worthy of publication.

However, I recommend revisions for the following points:

1). There should be a bit broader introduction with references and text that puts the issues observed in Lake Erie into global perspective. What other large lakes worldwide have this problem? What other solutions have these regions come up with. Would this study apply to these lakes?

2) Please clarify that you are using all in situ data in your study. I know the great lakes have remote sensing data as well. Why was this not also utilized?

3) Performance metrics – are these the commonly accepted metrics for this type of analysis? Please elaborate on why chosen.

4). In my view the MAE ane RMSE error are fairly high for the Chl-a model as discussed in this text:

“Overall, Chl-a concentrations predicted by process-based model GLM AED were reasonably accurate, even though disparities still occurred between predictions and 200 observations which may be due to the methods of sampling and the simplification of the complicated 201 ecological processes, such as grazing [15]”

Looking at figure 4, the model does not look to do that good of job to me. Please elaborate on what might be physically lacking in the model and decrease your wording of the model being “reasonably accurate” beyond the discussion of method of sampling or the grazing or other processing – or at least include more discussion of potential reasons for model disparities and explain why you find these results to be adequately good. The temperature results are really good but the model Chl-a look marginal in my opinion, but I don’t have a lot of experience with Chl-a modeling so please convince me through comparing with other modeling studies or discussion why these results are good enough…what is the general quality of modeling results in other Chl-a studies?

I find the discussion of “Prediction of future water quality” to be too short starting on line 273. Please expand on this section and give more details.

Please include a few sentences in the last section on limitation and suggested future work.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Chapter 1.

1. Why RMSE and MAE were chosen as Network Performance Indicators. Quite often there are publications describing for which data types and their ranges a given indicator should be used. They are usually given when looking for the best network structure.

Chapter 2.

2. Why 40 neurons were selected. I believe that if we choose MLP for analysis, overfitting, underfitting, robustness analysis should be performed. As it is currently done in publications, e.g. from the MDPI publishing house: According to me, when it comes to overfitting analysis, underfitting is enough, for example, to google:

metodology of overfitting and underfitting study of the rxamined neural network structures mdpi,

The same goes for robustness analysis:

robustness study of the examined neural network structures mdpi.

3. I believe that referring to the formula (1) from the publication [42] from 1993 is quite a risky procedure. In recent years, there has been a tremendous increase in knowledge of neural networks and capabilities. According to me, in 1993 the knowledge of neural networks was at a significantly different level than today.
4. Was the hyperparameter optimization performed prior to the network learning? The result of such a study was not presented. According to Me, it is currently performed commonly and perhaps that is why the authors did not present it. Please show the network learning hyperparameters along with the description of the optimization results.
   
   
5. Please include the Learning Parameter for the network learning process in the publication.
   
   
6. For Figures 4 and 5, I recommend presenting a histogram of neural network errors.

Summary:

The article is good.

I believe that minor corrections should be presented and made.

Good Job!

Best wishes and good luck !

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have done an excellent job responding to my comments and the paper is in great shape and ready for publication.