Machine-Learning Approach to Non-Destructive Biomass and Relative Growth Rate Estimation in Aeroponic Cultivation

Round 1

Reviewer 1 Report

Dear,

 

The review article clearly addresses the importance of using models for prediction in agriculture. The authors concluded by stating that artificial neural networks are the most used and recommended to estimate parameters in agriculture.

 

I suggest that the authors enrich the manuscript, using figures to exemplify this modeling. It will be of better understanding for readers and future researchers who will be based on the use of the models discussed in the work. I have not placed a note within the text, however, I would like to revise this manuscript again with the figures inserted.

Author Response

1. I suggest that the authors enrich the manuscript, using figures to exemplify this modeling. It will be of better understanding for readers and future researchers who will be based on the use of the models discussed in the work. I have not placed a note within the text, however, I would like to revise this manuscript again with the figures inserted
RESPONSE:
Thank you for the recommendation! We have now included Figure 1 (page 4) showing the experiment setup, Figure 2 (page 5) showing the image processing used in segmentation, Figure 3 (page 6) showing the effect of log-transform and normalization on biomass, and Figure 5 (page 7) showing the network architecture used. The content explained by these figures were requested by the other referees as well.

Reviewer 2 Report

The authors compare the performance of a multi-variate regression network and a ResNet-50-based neural network in forecasting plant biomass as well as the relative growth rate from a short sequence of temporal images from plants in aeroponic cultivation. 

This is not surprising that hot machine learning technology can be applied in this field. Interestingly, the results of the author's research prove the differences between the two kinds of machine learning methods used and guide their potential applications. However, I suggest that the author pay more attention to the differences and commonalities between the two technologies, which is very important, because there are many advanced networks at present, and we can't test them one by one, but the development of advanced networks is inheritable.

Another area I focus on is the construction of the training dataset. Do the change in soil moisture and other factors affect the acquisition of true values?

Author Response

 1. This is not surprising that hot machine learning technology can be applied in this field. Interestingly, the results of the author's research prove the differences between the two kinds of machine learning methods used and guide their potential applications. However, I suggest that the author pay more attention to the differences and commonalities between the two technologies, which is very important, because there are many advanced networks at present, and we can't test them one by one, but the development of advanced networks is inheritable.
RESPONSE: 
We agree with the Referee that indeed we cannot test advanced networks one by one. Instead in this work we put forth a procedure which uses such a neural network approach to automatically and non-destructively ascert RGR and biomass. So in effect the two different approaches tested here are not so important in terms of our overall message which is that using such technologies is better than the classic approaches which measure RGR and biomass by removing leaves from plants (destructively). In fact we now clarify this point in the manuscript (in lines 62-69). 

2. Another area I focus on is the construction of the training dataset. Do the change in soil moisture and other factors affect the acquisition of true values?
RESPONSE: 
In general for any type of plant growing environment as long as data is recorded consistently the resulting model will be able to learn. In our study we only require a short series of images in order to infer RGR and biomass. In that respect changes in atmospheric or soil moisture did not influence at all the imaging of the plants which in turn are responsible for the accuracy of our estimates. 
We now include these comments in the manuscript in lines 335-339.
In addition, to address the Referee question about the construction of the data set we now provide more such details through the addition of Figures 1, 2 and 3 on pages 4, 5, and 6 respectively. These figures highlight the experiment setup and process in higher detail, which was also requested by the other Referees. In addition, a new section has been added (section 3.1 experiment setup) which provides further details.

Reviewer 3 Report

I have gone through the article “Machine learning approach to Non-Destructive Biomass and Relative Growth Rate Estimation in Aeroponic Cultivation”. The topic has immense importance in the field of agriculture, where meeting the food security is a challenge with ever increasing population and adverse effect of climate change. The manuscript is nicely structured with well-defined research gaps. However, I found that there is scope to improve the manuscript. Hence, I would recommend a MAJOR REVISION for the article in its present form, before it is accepted for publication. The details review is provided below:

 

Introduction:

·        The initial two paragraphs require reference in support of these statements.

·        Reference may be added for statement (line no. 68-69)

 

Data processing and Methods:

·        The authors have provided elaborate descriptions of the methods and datasets. However, there were number of treatments in terms of plants, days, angle, transformation etc., it is suggested to provide a table or details of experimental design.

·        The authors are requested to provide an image of the experiment for better appraisal.

·        For better comparison, the scale of both the image in Figure 2 may be maintained.

 

Results:

·        The results are merely descriptions of the output. The insights for the varying results under different treatments will definitely add value to the manuscript.

·        The RMSE values are provided, however the authors must provide some curve or table of the absolute values of the biomass and RCR for better evaluation of the obtained results.

·        In line 265-267, authors have mentioned that “a poor estimate of biomass did not necessarily lead to a poor RGR estimate, even on an individual level”. The physical justification of the same may be provided, i.e., the better performance of (i) top camera in biomass estimation and (ii) dual mode in RGR may be provided. In similar line, the impacts of top, angle and dual camera on MVR and ResNet-50 may also be discussed.

·        The term “performed worse” may be replaced with “underperform”

 

Conclusions:

·        The major assumption of the study is the constant RGR. In that context, can the present methodology can really forecast the plant biomass and RGR, in real-life condition?

·        The second paragraph is redundant.

·        The conclusions may not utter the findings of the experiment, but the lesson learnt, significant value addition and way forward.

Comments for author File: Comments.docx

Author Response

Introduction:
1. The initial two paragraphs require reference in support of these statements.
RESPONSE:

We have now added references [1, 2, 3, 4] in order to better support the statements made in the initial two paragraphs.

2. Reference may be added for statement (line no. 68-69)
RESPONSE:

Indeed the sentence structure in these lines was a little unclear. The point we were trying to get across was that if destructive measurements are taken of the plants, the growth will be hindered or stopped. Thus, non-destructive measurements are necessary for gathering growth information in real time. We have now re-written these lines as follows: "Measurements based on removing parts of the plant can impact or even stop growth. Non-destructive machine learning approaches such as those we propose here are therefore needed." we have now added this text in lines 65 - 69.

 
Data processing and Methods:

3. The authors have provided elaborate descriptions of the methods and datasets. However, there were number of treatments in terms of plants, days, angle, transformation etc., it is suggested to provide a table or details of experimental design. DONE. 4. The authors are requested to provide an image of the experiment for better appraisal.
RESPONSE:

We have added a new section (3.1 Experiment Setup) which highlights the details of the aeroponic rig and treatments used for the experiments. We have also added Figure 1 in that section which shows the camera positions and growth bed design. In addition, Figure 2 has now been added on page 5, showing an example of the transformation used to extract and segment the images.

5. For better comparison, the scale of both the image in Figure 2 may be maintained.
RESPONSE:

We have updated the figure (now it is Figure 6) so that both figures have the same scale.
 
Results:
6. The results are merely descriptions of the output. The insights for the varying results under different treatments will definitely add value to the manuscript.
RESPONSE:

Indeed, we discuss the details and insights from these results in the Discussion section. Following the Referee recommendation we also provide further insights in the Conclusions section. If the Referee would like us to put together the Results and Discussion section we could do that instead.

7. The RMSE values are provided, however the authors must provide some curve or table of the absolute values of the biomass and RCR for better evaluation of the obtained results.
RESPONSE:

Our interpretation is that the Referee would like that the true biomass and RGR values should be presented along with the RMSE so that the 'relative RMSE' can be better evaluated. This has now been added to the Results section (lines 268-269) in order to allow for easier interpretation of the result.

8. In line 265-267, authors have mentioned that “a poor estimate of biomass did not necessarily lead to a poor RGR estimate, even on an individual level”. The physical justification of the same may be provided, i.e., the better performance of (i) top camera in biomass estimation and (ii) dual mode in RGR may be provided. In similar line, the impacts of top, angle and dual camera on MVR and ResNet-50 may also be discussed.
RESPONSE:

Indeed, this was not explained. We have now added the following explanation in lines 298-301: "The selection of the three random samples could on the other hand be affected by outliers. This might cause the variance-reducing effect of the dual view to be more prominent in the RGR estimate than in the biomass estimate, thus leading to an improvement in the former, but not the latter. "

9. The term “performed worse” may be replaced with “underperform”
RESPONSE:

We have now changed "performed worse" to "underperformed" in line 269 as suggested.
 
Conclusions:
10. The major assumption of the study is the constant RGR. In that context, can the present methodology can really forecast the plant biomass and RGR, in real-life condition?
RESPONSE.
We agree that this was indeed a bit unclear. The assumption of a constant RGR is only applied when constructing the data set and enables the high-frequency measurements used. This assumption is however not applied to the biomass models themselves. So non-constant RGR is not necessary for future data for the model (although for controlled hydroculture setups where the conditions are constant, this will likely be the case anyway). To clear this up, we have added a section explaining this on lines 160-163.
However, it is indeed correct that when we are comparing the models through the RGR estimates they provide, we are applying this assumption of the data set. As these estimates are made using the entire time span, the estimates can be considered to be the 'average' RGR, which is the same as the RGR if it is constant. A 5-day 'window' over the average is not that long, but if the RGR varies much over time, this could indeed affect the RGR estimates. However, since the method for estimating RGR depends on the plant, setup, conditions etc, there is no general solution to this. To explain this, we have added a paragraph explaining this in section 5.4 in lines 355-361: "It should, however, be noted that since the RGR estimate uses time points from along the entire growth period, the resulting estimate could be considered to measure the 'average' RGR. Since our data is assumed to have a constant RGR this does not matter. However, for conditions where the RGR varies over time (for example, if the conditions change during growth), the RGR should instead be constructed from a time series with a shorter time span depending on the time resolution desired. 

11. The second paragraph is redundant.
RESPONSE.
We have now removed the 2nd paragraph from the Conclusions as suggested.

12. The conclusions may not utter the findings of the experiment, but the lesson learnt, significant value addition and way forward.
RESPONSE.
Thank you! We have now removed the findings of the experiment from the conclusions and focused on the lessons learned, significant values addition and way forward as suggested by the Referee.

Round 2

Reviewer 1 Report

Dear,

The authors made the modifications as per the suggestions.