Using Machine Learning for Nutrient Content Detection of Aquaponics-Grown Plants Based on Spectral Data

Round 1

Reviewer 1 Report

To have a sustainable agriculture using aquaponics grown plant, it is very important to have accurate nutrient content information. In this purpose, your paper is very impressive and important. However, I have several question and comments as follow:

1)    Do you only study the fixed environmental condition (temperature, moisture, etc.)?

2)    I think the distance between net cups is also important to have growth condition information of lettuce. Do you check reflectance for different net cups distance?

3)    In 3.1, do you have any description of nutrient systems’ difference? I could not find any information about that. Without that, there is difficult to understand the difference among systems. Especially between system B and C.

4)    In figure 6, it seems that lower than 1150 um, there are difference among systems but in 3.3, you suddenly propose to use higher wavelengths too. Can you explain to fill in the gap between Ch 3.2 and 3.3?

5)    In (b) and (d) in figure 8, for me, there are big variance and there is no good linearity correlation between observed and fitted response. Can you justify it?

Author Response

Independent Review Report, Reviewer #1

Comments and Suggestions for Authors

To have a sustainable agriculture using aquaponics grown plant, it is very important to have accurate nutrient content information. In this purpose, your paper is very impressive and important. However, I have several question and comments as follow:

Authors: Thank you for recognizing the contribution of our work. We appreciate all your valuable comments and suggestions that enriched the article.

 

1. Do you only study the fixed environmental condition (temperature, moisture, etc.)?

Authors: Aquaponics systems are set up inside greenhouses, so there is some control over environmental parameters. But the proposed model can also be efficiently applied in the open field.

 

2. I think the distance between net cups is also important to have growth condition information of lettuce. Do you check reflectance for different net cups distance?

Authors: We are so sorry, we did not take into consideration the distance between net cups. We will consider it in future studies.

3. In 3.1, do you have any description of nutrient systems’ difference? I could not find any information about that. Without that, there is difficult to understand the difference among systems. Especially between system B and C.

Authors: Details of the nutrient systems used are in Section 2.2. Experimental Setup, page 3, lines 141 - 146.

 

4. In figure 6, it seems that lower than 1150 um, there are difference among systems but in 3.3, you suddenly propose to use higher wavelengths too. Can you explain to fill in the gap between Ch 3.2 and 3.3?

Authors: Thank you very much for this very valuable note. The wavelengths mentioned in Section 3.2 are the most affected by the plant's nutrient content to show the change in reflectance according to the plant's nutritional content. This does not mean that there are no other wavelengths affected by nutrients, it has been reported that different wavelengths are affected by the plant's nutrient content, which cannot be neglected. We have redressed this and included it in the text of the paper by citing the corresponding references, page 8, lines 294 to 299, and lines 315 to 318.

 

5. In (b) and (d) in figure 8, for me, there are big variance and there is no good linearity correlation between observed and fitted response. Can you justify it?

Authors: We appreciate this important comment. The models that did not perform well used the outputs of the PCA algorithm as the inputs. The features (components) selected by PCA are very few compared to those selected by SFS and GA. Due to the low number of features used in training the models, the linear correlation between the observed and fitted response is very weak. Therefore, the prediction accuracy was very low in cases of figure 8 (b) and (d).

Reviewer 2 Report

Line 25: It would be better to provide the name of the model for the spectroradiometer used in the study.

Lines 161-162: Apart from time periods, it would be interesting if you specify phenological phases of the crop when you conducted the measurements, because it is unclear whether the crop was in the same phenological phase each time or not. I suppose that lettuce was in the phase of technological ripeness, however, it would be better if you clarify this issue so that the readers don’t need to guess.

Line 259 and further in the same cases: R2 should have “2” in superscripts.

The results of the study prove significantly higher accuracy of BPNN and RF models. However, it is somewhat unclear how these models could be utilized in practice if we cannot get a clear equation as a result of modeling (and in the case of PLSR we obtain an equation). This point needs explanation in the discussion and/or conclusions section.

Author Response

Independent Review Report, Reviewer #2

 

Authors: Thank you for recognizing the contribution of our work. We appreciate all your valuable comments and suggestions that enriched the article.

 

Comments and Suggestions for Authors

1. Line 25: It would be better to provide the name of the model for the spectroradiometer used in the study.

Authors: Already the name of the model of the spectroradiometer has been provided, lines 25-26, thank you very much.

 

2. Lines 161-162: Apart from time periods, it would be interesting if you specify phenological phases of the crop when you conducted the measurements, because it is unclear whether the crop was in the same phenological phase each time or not. I suppose that lettuce was in the phase of technological ripeness, however, it would be better if you clarify this issue so that the readers don’t need to guess.

Authors: Thank you very much for this valuable comment. We have included the phenological phases along with the time period after transplanting the seedlings, page 4, lines 164-166.

 

3. Line 259 and further in the same cases: R2 should have “2” in superscripts.

Authors: Thanks for your consideration we already modified it in all manuscript.

 

4. The results of the study prove significantly higher accuracy of BPNN and RF models. However, it is somewhat unclear how these models could be utilized in practice if we cannot get a clear equation as a result of modeling (and in the case of PLSR we obtain an equation). This point needs explanation in the discussion and/or conclusions section.

Authors: We appreciate this comment very much, but it is difficult to extract an equation from models that have been trained to track nutrient content. After training the models and testing their predictive ability, a Python code (sav. file) is generated with the optimal hyperparameters to be used again in other practical practices (page 14, lines 433-436).

Round 2

Reviewer 1 Report

Please see the attachment.

Comments for author File: Comments.pdf

Author Response

Round 2, Independent Review Report, Reviewer #1

Dear Authors,
Thank you for your revision and reply. However, to reflect my comments, I would like to request you to add some text in the following comments.

Authors: We thank you very much for your valuable time to review our manuscript and we would like to express our gratitude for your comments that contribute to the enrichment of the manuscript. All your valuable comments have been noted and included in the manuscript

 

1. Do you only study the fixed environmental condition (temperature, moisture, etc.)?

Authors: We are grateful to have brought this to our attention, this comment has been taken into consideration and the manuscript text has been re-addressed on page 3, Section 2.2. Experimental Setup, Lines 140-145. The Conclusions section has also been re-addressed, lines 502-505.

 

2. I think the distance between net cups is also important to have growth condition information of lettuce. Do you check reflectance for different net cups distance?

Authors: We are so sorry, we did not take into consideration the distance between net cups. We will consider it in future studies. The effect of the distance between cups on reflectivity has been included in the Conclusions section as a future work lines 505-508.

 

3. In 3.1, do you have any description of nutrient systems’ difference? I could not find any information about that. Without that, there is difficult to understand the difference among systems. Especially between system B and C.

Authors: Details of the nutrient systems used are in Section 2.2. (Experimental Setup), page 3, lines 145-150. Also, details of nutritional systems are included on page 7, Section 3.1 (Changes in Lettuce Growth under Different Nutrient Levels), lines 277-282.

4. In figure 6, it seems that lower than 1150 um, there are difference among systems but in 3.3, you suddenly propose to use higher wavelengths too. Can you explain to fill in the gap between Ch 3.2 and 3.3?

Authors: Thank you very much for this very valuable note. The wavelengths mentioned in Section 3.2 are the most affected by the plant's nutrient content to show the change in reflectance according to the plant's nutritional content. This does not mean that there are no other wavelengths affected by nutrients, it has been reported that different wavelengths are affected by the plant's nutrient content, which cannot be neglected. We have redressed this and included it in the text of the paper by citing the corresponding references, page 8, lines 304 to 309, and lines 325-329.

 

5. In (b) and (d) in figure 8, for me, there are big variance and there is no good linearity correlation between observed and fitted response. Can you justify it?

Authors: We appreciate this important comment. The models that did not perform well used the outputs of the PCA algorithm as the inputs. The features (components) selected by PCA are very few compared to those selected by SFS and GA. Due to the low number of features used in training the models, the linear correlation between the observed and fitted response is very weak. Therefore, the prediction accuracy was very low in cases of figure 8 (b) and (d). Due to the importance of this comment, this interpretation has been included in the text of the manuscript on page 12, section 3.4. (Performance Evaluation of Regression Models), lines 438-443.

Author Response File: Author Response.pdf