Use of CNN for Water Stress Identification in Rice Fields Using Thermal Imagery

Round 1

Reviewer 1 Report

Summary

On the basis of thermal images of rice fields, the authors of the paper made use of a CNN (Convolutional Neural Network) in order to determine the level of water stress that the rice was experiencing.
They divided the irrigation situation into three categories: one with 100% irrigation, one with 90% irrigation, and one with 80% irrigation.
They first used CNN to extract from each thermal image a temperature level score, and then they used these scores to further classify the three different irrigation instances.

Overall, the paper is relevant,  well writing and contains sufficient scientific merit

Comments:

**I recommend that authors divide the introduction section into two sections: introduction and literature review.
The section on literature review should provide a "critical review" of previous research rather than a "descriptive" review. This critical review may include the limitations/comparisons/advantages/of previous research, as well as their classification (a classification table may be nice). For greater clarity, at the end of the literature review section, include the "gap" filled by the current work and clearly enumerate the paper's contributions.

** Regarding subsection 2.1 about Dataset and for results reproducity: The authors are encouraged to share such datasets on a public depository file and post the link in the paper (such as mendeley)

** In subsection 2.2, First, there is a confusing notation problem in the sentence: "For convenience, we use symbols to represent TLSef and TLSlf respectively, and then we use MLP to identity the final water stress situation classes according to these two scores" Please specify the meaning of TLSef and TLSlf
Second, the authors used MLP to identify the final water stress situation classes according to the temperature scores.
My question is: to prove the utility of the MLP, is it relevant to compare your current framework with a framework without MLP?
If the framework without MLP is less efficient then, the role and the use of MLP is justified.

** In subsection, The authors stated that: "The activation functions of these three layers are all ReLU." I think that all choices (activation functions, some values of parameters, etc) should be justified. I know that these functins are classical ones in machine learning literature, however, if, for instance, authors compare an architecture with Relu functions with another architecture with different functions and then conclude that the architecture with Relu is the best one, Then, the choice of Relu is (at less experimentally) justified

** Overall, the paper should be strengthened by adding more elaborate results (10 pages are not enough)

The authors are invited to check some typos in the text

Author Response

1-1. We use Table 1. to list the essentials of 6 related literature and give an appropriate critical review in Remark.
1-2. We also list this contribution on lines 130 to 139 of the modified version (including revision tracking)

2. Due to other researches on the related dataset, I am sorry that the dataset cannot be made public on the Internet at now, but the dataset can be provided by contacting the author, and we also indicate it at the end of the article.

3. We added the notes for both of two scores in the revised version on lines 198-199.

4. We have added related experiments, using two TLS directly on the classification, they are written in lines 317-333 of the revised version.

5. For the activation function of MLP, we added the experiment of tanh and sigmoid function, and related descriptions are in line 353 to line 373 of the revised version.

6. Our revised version has 11 pages.

7. We have edited English by native speakers and corrected spelling mistakes. Also, in order to keep the article clear, most of these English revisions are not tracked by revisions.

Thanks again to the reviewers for taking the time to review the manuscript.

Author Response File: Author Response.docx

Reviewer 2 Report

Li et al. evaluated the possibility of water deprivation identification based on thermal images and CNN. The data set is promising. Authors are requested to deal with the following comments within the manuscript (rather than the response letter).

 

(1) Authors used 3 irrigation regimes (100, 90 and 80%). However, they do not mention percentage of what? Is it 100, 90 and 80% of evapotranspiration?

(2) They need to motivate why they used these 3 irrigation regimes (100, 90 and 80%). The classification needs to be consistent (different in Lines 85–87 as compared to 108–110).

(3) Is there any yield penalty under reduced irrigation under study? Literature?

(4) which cultivar was used? Cultivation period and duration? Growth conditions (light intensity, temperature, relative air humidity) during plant growth?

(5) pictures were assessed in fully-grown plants?

(6) time of the day and distance to plant when obtaining the images

(7) you need to provide the spectral range, where images were obtained (e.g. 400–700 nm)

 

(8) Lines 29–30 (introduction): Climate models project declining rainfall frequency allied with rising temperatures, which will further increase the gap between actual and potential yield (Seifikalhor et al. 2022 Scientific Reports 12, 7034). In other words, authors may stress that existing yield penalty is expected to increase in the near future owing to climate change effects.

(9) Lines 52–54 (introduction): In the last decade, CNNs have been increasingly employed in plant phenotyping community. They have been very effective in modeling complicated concepts, owing to their ability of distinguishing patterns and extracting regularities from data. Examples include variety identification in seeds (Taheri-Garavand et al., 2021 Plants 10, 1406) and in intact plants by using leaves (Nasiri et al., 2021 Plants 10, 1628).

(10) (introduction) Indeed estimating leaf water content by employing cameras operating in the visible portion of the electromagnetic spectrum (400–700 nm) compensates the cost factor (Taheri-Garavand et al., 2021 Acta Physiologia Plantarum 43, 78). Such methodology, however, not only requires positioning of the leaf at a specific orientation in relation to the camera, but also defined illumination conditions (Taheri-Garavand et al., 2021 Acta Physiologia Plantarum 43, 78). The required specific illumination limits the method applicability to controlled-light environments. Is thermal imaging a method independent of the ambient light environment? Is this solution affordable/cost-effective, portable? Does and it provide rapid measurements.

 

(11) what computation power, time and resources are needed to run such an algorithm?

(12) who are the intended users? Farmers, scientists, students?

(13) is this real-time or is it intended to become?

(14) can the present algorithm applied in other species? Or the same process ought to be performed again?

(15) Enrich the discussion by including: ”On a commercial scale, evidently, a capital investment is initially required for adopting the employed approach (Taheri-Garavand et al., 2021 Industrial Crop Prod 171, 113985). Nevertheless, the wide-ranging large-scale commercial applications can provide high returns through considerable improvements in process enhancement and cost reduction.”

(16) The discussion is very poor and ought to be improved. You need to compared your findings with previously-obtained ones, and highlight what is new and novel

(17) A single rice cultivar was used. Do you expect that the findings can be readily reflected in other rice cultivars too?

moderate editing is needed

Author Response

(1) We use reference 7 to illustrate this experimental setting, their watering rate is 100%-50%, and our experiment is 100%-80% when the water shortage is not serious. Relevant instructions are in line 165 to line 171 of the revised version(including revision tracking).

(2)We have made them consistent, they are in the revised version of lines 117 to 119 and lines 170 to 171 respectively.

(3)For yield penalty, we add references 4-7 to illustrate it, they are in revised version 39 to 47 lines. 

(4)I'm sorry that I didn't record these data during the experiment, only the date of planting is described in lines 178 to 179 of the revised version.

(5)Yes, they are fully-grown plants.

(6)They are described in line 179 of the revised version.

(7)They are described in line 181 of the revised version.

(8)We use this reference [12] to illustrate this drought warning on other crops.

(9)We have briefly described these literatures in lines 65 to 68 of the revised version.

(10-1)We have included this reference in the [27] and described their major shortcomings in the revised version of lines 73 to 78, and Table 1. 

(10-2)For the cost issue, we recommend using an external thermometer connected to the mobile phone. This cost is low, and we mentioned it at the end of the conclusion (lines 395-396 of the revised version).

(11) In fact, we use 2080ti GPU for neural network calculations, but in this article, it is real-time to predict each data, and it is not meant to list the training time, so we did not include the description of computing power in the article. And The practical application scenario suggestion is described on line 395 of the revised.

(12)Because it is low-cost, we consider it unrestricted to the user group, which we describe briefly in the conclusion (lines 395-396 of the modified version).

(13)For predicted, it is real-time.

(14)For other varieties or plants, it must be retrained, because their phototype gap is usually not small, but retraining this framework is not a big project. We also give a brief note on line 138 of the revised version

(15)Commercialization is our goal. Our method is low-cost. For users, only need a mobile phone connected to an external thermal imager. For the server, such computation costs are not too high for multiple users.

(16)For references 27-32, the biggest problem is:
1. Differences between RGB and thermal images
2. Consideration of lack of degree information

We have added several experiments in the discussion section to improve the explanation that we have made up for the deficiencies of previous studies.

For example, Table 4 and 5 for 1. Table 10 for 1. 

(17)We conduct experiments on three rice varieties Tainan 11, HVA1, and TNG67 (modified version 165), but we really need to ink on variety generalization for the future.

We have edited English by native speakers and corrected spelling mistakes. Also, in order to keep the article clear, most of these English revisions are not tracked by revisions.

Thanks again to the reviewers for taking the time to review the manuscript.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear authors,

** Please modify the title to "Use of CNN for Water Stress Identification in Rice Field with Thermal Images"

** In keywords: avoid the abbreviation CNN

 

Reviewer 2 Report

Authors did excellent work in dealing with my comments. The quality of the manuscript has been significantly improved.

I, now, recommend the manuscript for publication.

minor changes needed