In this paper, the authors use a deep neural network to predict awn phenotype and percentage heading in the wheat. Although the paper is written in a good English, the overall organisation and methodology is rather hard to follow.

From the CNN perspective, it is unclear what the input is (a cropped image for my understanding), and it is even more unclear what the outputs are. In one case, they train to predict the awn phenotype, but what does it mean? Is the network trained to find awned and awnless genotype (binary classification?). When they predict the percentage heading, is this another network? Moreover, why they use classification, instead of regression, to predict the percentage heading? I believe continuous number prediction might fit better in this context.

As a follow up, in line 89, what are these breeders score the authors refers to?

By looking at the results of awn and awnless plants, an accuracy of the ~98% indicates that the problem is rather easy for the DNN. My experience in computer vision, in particular in plant phenotyping, has taught me to use DNN when the problem you are trying to address is very hard, due to huge intra and/or inter class variability. This is also clear in a recent publication by Pound et al. 2017. They have a DNN performing two tasks and the one classifying the awn plants reaches 98%. But, in that case, they also ask to the network to localise the wheat spikes.

This also rises another point in this paper. I think the authors have not performed a diligent work in their literature review. The paper I mentioned is very relevant to their problem. However, there is a broad literature of DNN and plant phenotyping problems (leaf count, leaf segmentation, multi-modal images, etc.). As an example, Dobrescu et al (2017) and Giuffrida et al (2018) use ResNet, which is the same architecture the authors use in this paper. At line 320, the authors say Resnet has been used in many applications, ignoring completely all the work that has been done in plant phenotyping.

Personally, I found hard to see how DNN links with genetic analysis. I think the connection between gene and DNN output must be better explained.

At line 91, are they thinking to release the annotated images used for this paper? Because I think this is very important and if the authors do that, it would be very beneficial for the plant community.

Overall, I think the paper has merit, however it needs a lot of work to warrant a publication in Gigascience. All the arguments need to be rearranged and better explained, linking them from DNN results to genetic analysis. I also suggest to reconsider what image is supplemental and what should go in the main manuscript. Only two images are in the main manuscript, and all the rest (included tables) are as supplemental. I found this choice not very optimal.

References: Pound et al. (2017): Deep learning for multi-task plant phenotyping Dobrescu et al. (2017): Leveraging multiple datasets for deep leaf counting Giuffrida et al. (2018): Pheno‐Deep Counter: a unified and versatile deep learning architecture for leaf counting

Declaration of competing interests Please complete a declaration of competing interests, considering the following questions: Have you in the past five years received reimbursements, fees, funding, or salary from an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold any stocks or shares in an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold or are you currently applying for any patents relating to the content of the manuscript? Have you received reimbursements, fees, funding, or salary from an organization that holds or has applied for patents relating to the content of the manuscript? Do you have any other financial competing interests? Do you have any non-financial competing interests in relation to this paper? If you can answer no to all of the above, write 'I declare that I have no competing interests' below. If your reply is yes to any, please give details below.
I declare that I have no competing interests.

I agree to the open peer review policy of the journal. I understand that my name will be included on my report to the authors and, if the manuscript is accepted for publication, my named report including any attachments I upload will be posted on the website along with the authors' responses. I agree for my report to be made available under an Open Access Creative Commons CC-BY license (http://creativecommons.org/licenses/by/4.0/). I understand that any comments which I do not wish to be included in my named report can be included as confidential comments to the editors, which will not be published.
I agree to the open peer review policy of the journal.

Authors' response to reviews: (https://drive.google.com/open?id=1TYyOraQQJu18xGUajSmnj9qtXsEol30K)

The paper Wang et al. presents a DNN to extract phenotype from Wheat plants. The paper has improved from its first revision and it has nice experiments merging phenotype and genotype. I think the paper should be accepted after a few minor edits (which I highlight below). It is of great interest the authors will release the data with the paper, which enhance the value of this research for the broad community. I appreciate the efforts of the authors to release data and annotations. Moreover, what I found interesting is the extraction of phenotype traits in the field.

My minor comments are:

Line 39-42 "Most powerfully, the deep learning approach presented here gives a conceptual advancement in high-throughput plant phenotyping as it can potentially estimate any trait in any plant species through leveraging expert knowledge from breeders, geneticist, pathologists and physiologists." I think this is a rather bold statement. Supposing that I want to predict root mass from the shoot images and suppose that I have annotated >10,000 images/plants, where I manually harvested them and took the root (dry) mass. Do the author believe it is possible to learn a DNN to create such a mapping? Of course it can be done, but I am rather skeptic about the robustness of the predictions.

Line 62-64: DL does not discovers and end-to-end process. DL optimises network parameters in an end-to-end training (most of the times). The way it is written, it seems to me that somehow DL optimises for the network architecture. Line 66: I would say "the first filters are easily interpreted as LOW LEVEL image features"

Line 434: ResNet (or any dnn) is not restricted in the size of its input. It actually the opposite. The minimum image size resent accepts is ~300x300 (more or less). Ideally, one can provide any image size as input. The problem comes for the fully connected layers that will have millions of parameters with such a big images. Therefore, such a big network will saturate the memory of the gpu, not allowing the training. The way it is written sounds like it is a limitation of the architecture, but it is actually a (current) limitation of the hardware.

In general, I think the authors should tone down how impressive DL/CNNs. DL has been introduced in the Plant Phenotyping community since 2016. I think all the people involved in this community are aware of how DL is great in many context.

Ideas how to make this paper even better (these suggestions are not mandatory for this submission, but the authors are encouraged to follow them) The authors make a claim from line 446 about visual scoring discrepancy. This is basically known in the literature as inter-observer (or inter-rater) variability. The authors might perform a nice study about that and check whether the machine predictions are within the human variability. I still believe the scoring predictions should be performed as a regression task, rather than a classification task.

Declaration of competing interests Please complete a declaration of competing interests, considering the following questions: Have you in the past five years received reimbursements, fees, funding, or salary from an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold any stocks or shares in an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold or are you currently applying for any patents relating to the content of the manuscript? Have you received reimbursements, fees, funding, or salary from an organization that holds or has applied for patents relating to the content of the manuscript? Do you have any other financial competing interests? Do you have any non-financial competing interests in relation to this paper? If you can answer no to all of the above, write 'I declare that I have no competing interests' below. If your reply is yes to any, please give details below.
I declare that I have no competing interests.

I agree to the open peer review policy of the journal. I understand that my name will be included on my report to the authors and, if the manuscript is accepted for publication, my named report including any attachments I upload will be posted on the website along with the authors' responses. I agree for my report to be made available under an Open Access Creative Commons CC-BY license (http://creativecommons.org/licenses/by/4.0/). I understand that any comments which I do not wish to be included in my named report can be included as confidential comments to the editors, which will not be published.
I agree to the open peer review policy of the journal.

Authors' response to reviews: (https://drive.google.com/open?id=1gkH5wBKkcknT_F5wXj7IhF2KhSxr4gfg)