Artificial Intelligence and Novel Sensing Technologies for Assessing Downy Mildew in Grapevine

Round 1

Reviewer 1 Report

• Brief manuscript summary

The manuscript examines the opportunity to adopt novel non-invasive sensing technologies, concerning computer vision, hyperspectral imaging and artificial intelligence, for monitoring grapevine downy mildew, in terms of disease detection, early detection and severity disease under laboratory and field condition. Under laboratory condition, leaf discs from a susceptible to Plasmopara viticola cultivar were obtained: one group was inoculated with the pathogen, the other group was used as control. The images acquired for nine days after inoculation with a RGB camera were subsequently processed to quantify leaf discs area covered by sporulation and to assessing severity disease. The results were compared with the evaluation carried out by eight expert panelist. The RGB images obtained in field were processed to detect downy mildew symptoms. Moreover, hyperspectral imaging was used for the early detection of infection.

The authors conclude that these technologies could be promising for detecting grapevine downy mildew.

• Broad comments

The aim of this work is very interesting: the possibility to adopt novel sensing technologies for detecting grapevine downy mildew and for evaluating the level of disease severity gives the opportunity to improve speed and accuracy of disease assessment.

While I understand the experimental methods carried out under laboratory condition, the analysis concerning field condition are quite confusing and the results are not well discussed.

Suggestion: the sporulation is not a symptom so, in the manuscript when authors refer to sporulation, replace the term “symptom(s)”.

• Specific comments

L46 - 47.  The evaluation of grapevine downy mildew is based on visual assessment in both laboratory and field. The histological analysis are not carried out in order to assess the disease, but to analyse other factors (such as the response of a resistant plant to the pathogen).

L77. It is not mentioned the name of the cultivar.

L80 - 81. This part is not clear: why the authors refer to “leaf discs under field conditions in vineyard”?

L84. Replace the capital letter “Different” with “different”

L115. “Finally” is not appropriate in this context.

L155-158. Where are the results obtained analysing the control leaf discs in Figure 5?

L 163. How the authors obtained this data? How was calculated the accuracy percentage?

Author Response

After reading the reviewers’ comments, we have carefully modified and improved the work based on their suggestions. We first explain the general changes and then we provide a more detailed explanation to the reviewers’ comments.

The main changes introduced in our new manuscript are the following:

• All new text content (modifications, new sentences, data, etc.) was highlighted in green in the document to facilitate its tracking.
• Figure 1 has been modified.
• New table 1 has been added.

• All text was carefully reviewed and improved for better readability.

Reviewer 2 Report

The paper addresses the problem of downy mildew assessment using non-invasive methodologies.

Some general comments:

It seems that the work is clearly divided in two different learning approaches: a non-Deep Learning and a Deep Learning one, but the reason to present both (since the DL based exhibits better results) is not clear.

Do you think that different grape varieties can lead to different downy mildew conditions (with impact in the visible leaf characteristics and /or in the spectral information) and, as such, the model used has to be adapted to every variety?

The conclusions are short and could be further developed.

Specific comments:

2.1

It is not clear if the hyperspectral images were taken under field conditions, because they were taken on leaf disks.

2.2

It is not clear how GrabCut was used, since it usually needs user input.

Why were the pre-processing techniques used?

The reason underlying the color space may be improved. The human perception in these models does not seem to be the only reason, but also the non-correlation of the 3 channels which turn HSV or HLS more suitable for some of the tasks.

The use of global thresholding (Otsu) needs to be clarified. Was is to segment downy mildew areas in the leaf disks? If so, why didn’t you perform DL semantic segmentation on the leaf disks images

2.3

The CNN model description for hyperspectral images is too vague since it may correspond to many different DL architectures.

3.

Do the experts visually assessed the extent of downy mildew delineating the affected area or do they did it in some other way? If so, what is the coefficient of determination referring to?

A table with results for the different learning techniques used would be helpful to follow the discussion.

Author Response

After reading the reviewers’ comments, we have carefully modified and improved the work based on their suggestions. We first explain the general changes and then we provide a more detailed explanation to the reviewers’ comments.

The main changes introduced in our new manuscript are the following:

• All new text content (modifications, new sentences, data, etc.) was highlighted in green in the document to facilitate its tracking.
• Figure 1 has been modified.
• New table 1 has been added.

• All text was carefully reviewed and improved for better readability.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have improved the paper, namely restricting the study to laboratory conditions. Nevertheless, some aspects need attention:

Although some changes were introduced addressing the use of two different learning approaches – one for disease severity evaluation, the other for early classification-, it is my opinion that some of the arguments presented on the authors’ response (cost and availability/easiness of use) should be introduced in the text. I also believe that disease severity estimation might be a harder problem to deal with DL (if the severity is expressed as a percentage, we are talking about a regression problem or at least of a many classes classification one).

The impact of grape varieties in the methodologies analysis presented in the authors’ response should be (partially) included in the paper, for instance in the discussion and maybe in the conclusions – as a possible limitation.

The new table 1 is very useful and clear but when discussing computational cost (lines 141-143), and for trainable methods, the difference between training and testing time is important, since most of the times training is seldom performed compared to testing. So, the authors should clearly state what they refer to when they say the models take up to one hour when using CPU processing.

It should also be clear what the “modelling for the pathogen detection” (line 125) does: is it a binary classifier (presence/no presence of the disease)?

Author Response

The reponse to reviewer is included

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

The authors have addressed all comments and haver changed the paper accordingly.

I have no further comments.