Spatiotemporal Deep Learning Model for Prediction of Taif Rose Phenotyping

Round 1

Reviewer 1 Report

The authors presented deep neural networks to predict different phenotypes of Taif rose using different spatiotemporal dataset. I believe the authors have failed to provide enough scientific evidence behind the acceptance of the models presented. Therefore, I cannot recommend the current manuscript for acceptance without doing major corrections. Here I will first point out the most important concerns about the methods and results:

• I have found the total sample size of the proposed model is 10 samples. If that is the case, I cannot recommend this work to be published. Because deep neural networks usually require higher number of sample size. I ask the authors to clarify the sample size more clearly.
• The authors have provided the justification for using LSTM in this work at some extent, but data-driven models are highly biased towards the availability of training data. Therefore, in the remote sensing and plant phenotyping community, it is a common practice to apply some novel data-driven technique, and then compare the result with other traditional techniques. For instance, the authors could apply other traditional machine learning models (i.e., support vector regression, partial least squares regression, random forest regression, or even simply the multiple linear regression) and compare the results they got from LSTM with the traditional ones. This would justify the performance of the proposed method.
• The authors have evaluated the model performance by showing MAE, MAPE, MSE, MSPE. Although, I would argue that most of the plant phenotyping literature show MSE, RMSE and may be normalized RMSE as evaluation metrics. However, the authors created a model which predicts 15 different phenotypes at once and one loss function takes care of the backpropagation from those 15 output variables. I doubt that the result might be better if they trained 15 different LSTM models for 15 different target variables. It would be interesting to see which way the model learns better (in terms of both performance and training time): learning all the 15 variables together or separately?
• Since deep neural networks can be overfitted easily, the common practice is to see the training loss curve and a validation loss curve for each epoch. The authors have provided 4 different errors but provided only training metrics. If the sample size is only 10, then it is hard to split the dataset into training, test and validation set. Therefore, without more sample size it is hard to say the deep neural networks performed without overfitting or underfitting. If the sample size is much higher than 10, then the authors should report the loss for both training and validation set.

Other than these major concerns, I also have some minor concerns. However, I believe a lot of writing has to be adjusted so I am not going to provide the full minor suggestions at this stage. Regardless of that, let me give some ideas about the minor suggestions.

• The overall writing of the manuscript needs to be improved. For instance, the abstract does not mention anything about the LSTM, or CNN or MLP. Both the methods and results are not clear in the abstract.
• Figures need to be improved. For instance, Figure 1 does not maintain the standards of map making. There are three different maps of three different scales, but no scale bar is present. The texts are hard to read and there are too much blank spaces without any information. I don’t think this map do no provide any useful information to the readers. Additionally, Figure 7-10 is not necessary to be shown as different. Rather, these can be shown in a single figure.
• There are unnecessary detailed tables in the manuscript. For instance, the statistical summary detailed in Table 2 is boring and cluttered with numbers. It will be much easier if this was a figure. I envision this type of descriptive statistics to be shown in histograms or box-whisker plots. Same comments for Table 5.

I thank the authors to conduct this interesting study. However, I also requests the authors to clarify my concerns.

Author Response

Authors wish to thank all reviewers for their effort and for the valuable suggestions they offered to us, and we hope that we addressed all of their major concerns to make this paper suitable for your journal. pls find the response is uploaded below

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

The author tried to determine new DL model applied to Taif Rose Breeding (TRB) combining satellite and phenotype, which can accelerate breeding efficiency. However, As the plot for plant breeding is small with a large number, it’s hard to precisely extract phenotypic traits due to the low spatial resolution of satellite images. So, it can't provide phenotypic traits for each plot combining the satellite data and sampled phenotypic traits at a large scale. Moreover, there also exist mixed pixel problem, which limit its application in the early stage. Major/Minor concerns are as followed.

There're too many background knowledge in the section of "Abstract", which lack of sufficient description for the methods and results. 

The section of "Related Works" should be reorganized. There're too many textbook knowledge, which lack of scientific summary. The author should deeply analyze the advantages and disadvantages of different methods for phenotyping, not only DL. There're many supervised and unsupervised methods that have been adopted in the study of plant phenotyping. Moreover, the empirical modeling method should also be included in the manuscript.

The dataset size in Table 1 should provide unit.

Line189-190, the statement is too subjective. There're many research that haven't been included in the literature review, so the author judged that "most of the related works design their DL models based on a single dataset that is either taken from in-situ or satellite imagery data sources". In fact, there're many research using multi-source data for the phenotyping in plant breeding.

Line 486-487, it's Phenotypes, not phonotypes.

The section of "Results and discussion "should be reorganized. Maybe divided into two parts.

Author Response

Authors wish to thank all reviewers for their effort and for the valuable suggestions they offered to us, and we hope that we addressed all of their major concerns to make this paper suitable for your journal. pls find the response is uploaded below

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Please see the attachment

Comments for author File: Comments.pdf

Author Response

Dear Professor Editor in Chief

 

We highly appreciate the time and efforts dedicated to reviewing our manuscript and the constructive comments and invaluable suggestions provided by you and the esteemed reviewers. These comments allow us to significantly improve and strengthen the content of our paper entitled " Spatiotemporal Deep learning Model for prediction of Taif Rose Phenotyping". Therefore, the all-out effort has been expended to address the comments and incorporate them in the revised manuscript accordingly. 

In the following, please find our responses to your comments where the reviewers' comments appear in black and our responses in blue. Two versions of the manuscript have been uploaded to Manuscript Central:

1. Old_MARKED3.docx includes track changes to highlight modifications to the original paper.
2. New_UNMARKED3.docx is the final version of the revised paper with all changes accepted.

 Pls. concern the number of mentioned lines in our responses in the version with tracking changes

Author Response File: Author Response.pdf

Reviewer 2 Report

All the comments have been addressed properly.

Author Response

we thank the Reviewer for his/her significant comments which help us to improve our manuscript efficiently