Review Reports

Round 1

Reviewer 1 Report

Dear Authors,

This manuscript describe the use of deep learning for the classification of lodging in sunflower. The authors used multispectral UAV imagery to characterize the crop and identify areas cultivated with sunflower, and patches affected by lodging within these areas. Despite the importance of the topic and the potential interest of the scientific community by the methods adopted in your study, several points need to be addressed before a possible publication. First the authors need to better describe how the multispectral data was used with the deep learning models, since these are originally designed for RGB images. In addition, the performance of the models was relatively poor but several avenues to improve this performance have not been investigated. For example, how the radiometric calibration of the images was made? If reflectance values cannot be derived and only raw data is available then the results of your study are compromised, especially if acquisition in the different fields was made under different illumination conditions (please read Section 4 of this article for refence 10.3390/rs10071091). Also, the authors did not try to explore the effects of different degrees of salinity or even of lodging to check if results can be improved. This is crucial to avoid misinform the reader, which might think that the approach does not work, but in fact the dataset used has not been well explored/prepared. Tackling these points is necessary in order to improve the material presented for a possible publication. Please find attached some specific recommendation for your manuscript.

Best regards,

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper applies two well-known semantic segmentation algorithms, namely SegNet and UNet using input from multispectral images captured from UAVs for the detection of areas with sunflower lodging in sunflower crops.
The two methods are seen to compare favourably with the Random Forest algorithm.
Actually, a variant of UNET algorithm was used, namely IEU-Net, from [38], which introduces  a new "ignore-edges categorical cross entropy function" as well as Dropout and Batch Normalization layers.
Unfortunately, [38] is avaliable only in Chinese language, and it a bit hard to understand all details from this paper.

This is clearly an application paper with almost no theoretical contributions and very minor innovations introduced. 
Also, results show that even the best perfoming combination (SegNet with input from RGB+NIR bands) has a lot of problems in detecting sunflower lodging(low recall), although precision is quite good.

However, the problem is studied quite extensively, many issues are discussed in depth and some interesting conclusions and directions for future research are provided.

The paper [30] is on exactly the same subject and presents very good results using an improved SegNet algorithm. Although it is briefly mentioned in the introduction, it is not analysed in more detail. 
So, I would recommend to the authors to add an additional section e.g. in the discussion section, comparing the techniques they used with those used in [30]. 

Specific comments:
l. 53 : "Kthe" => "the"
l. 256: Is there any justification or reference for the choice for this choice for mtry?
l. 239: "8 classification methods by combining data from four bands": "four bands" is not accurate: As you mention in l.236 you have 4 combinations of 3 to 5 bands.
l. 275: Finally, through softmax, the maximum value of different classifications is obtained as output, to get a segmented graph.
l. 321: "with the order of cutting" is unclear. Also the following sentence "The commonly used method of prediction result" needs to be revised.
l. 323-328: The description fo the method of ignoring the edges ([38]) is not clear enough and needs to be improved.
l. 352: Please define the loss function  and "accuracy" for both networks. Is it pixel-wise crossentropy loss? And are all 3-5 channels equally weighted?
l. 383: Please explain the difference between experiments in Sections 3.1 (Table 4) and 3.2 (Table 5). Initially, I thought it is one single experiment, but in 3.2 you focus on the results for the lodging class, but then I see "overall accuracies" that are different from those in Table 4. Please explain.
l. 434: Are the results in Table 5 for the validation set?
l. 457: I see no reason for including this Table in Discussion and not in the Results section. The same 
holds for Section 4.2. I propose to move detailed results to Section 3 and have a more concise Discussion 
section focusing only on the important remarks.
l. 480and l.490: Tables 1 and 2 => (Probably) Tables 5 and 6.
l. 540: The "traces of splicing" are not so "prominent" to an inexperienced reader. I suggest to somehow indicate these problems in the images, maybe along with the stiching of patches).

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Minors:
List of abbreviations needed at the end of the paper 

Example images - please add multispectral bands
Description of images (titles) should be with image also

Major:
Fig.8 shows accuracy and loss.
Extending numer of epoch is required - asymptotic convergence is visible, but convergence is not achived to stable value by limit of numer of epochs to 50.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

ok