Pre-Processing Training Data Improves Accuracy and Generalisability of Convolutional Neural Network Based Landscape Semantic Segmentation

Round 1

Reviewer 1 Report

1.How did you create the training data? Labeling method or process needs explanation.

2.Among the major classifications of land use classification, classes corresponding to crops were selected, but there is no explanation why only crop classes were selected.

3.The training data is 2018 and the test data is 2015. What season or month data was applied?

4.Does Number of patches mean number of samples? It is necessary to explain why the word patch is used instead of sample number.

5.Table 4. Number of patches  => It is common to make the number of learning data for each class uniform, but in the case of tree area and vineyard berry crop, there are relatively few problems.

6.Figure 14 => There seems to be little difference between the two pictures. It is necessary to indicate on the map which area is evident.

7. In title, "Pre-processing Training Data Improves Accuracy and Generalisability of Convolutional Neural Network" => This is a very common story, but does it really need to be used as a title? 3 page 119 "establish standard training data processing recommendations" => Shown as a more appropriate title.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Review form

Rows 133-134: You need to introduce as Figure 1 a map with the North Queensland district, also mark other important features for a better identification. The actual Figure 1 is very similar with Figure 2a.

Rows 163-170: Please provide a table with cover classes and indicate the number of polygons samples and pixels for each class

Rows 172-172. Please combine the two sentences

Figure 6. Not sure about this example, check once the credentials

Firs of all the methodology must be compress and shorten, please use formula or charts for some explanations, for example accuracy assessment.

I am not happy to put together the results and discussion please try and separate it, the results are the product and discussion are the interpretation of this new data and comparing yours results with others authors.

On the figure 14. Please brings same elements for orientation, rivers, major cities, roads, boundary of districts, or else you want

Figure 15. I sincerely appreciate it, looks nice and comprehensive

The article debates an interesting and actual topic “deep learning” technique used for a better classification of VHR images. The authors did a lot of works to customize the HPC resource, define training samples and classify the image. I appreciate it, the results look very good, but they still need to arrange the results and discuss it in a more concise way. I recommend a major revision of this paper with focus on sentences and phrases management.

English is good.Tthe problems are related to the sentences that must be more concise and the text compressed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

As I told you previously this small experiment has an actual topic with a huge potential for increase and is able to produce changes in the satellite data usage and processing.

The methodology chapter is not clearly defined and the workflow doesn’t t provide enough information regarding the steps followed. Also, a mixture of information could rather find its place in the discussion chapter.

I just know that it’s not simple, but the authors must make an effort to disseminate this information to other researchers in an easy way to make readers able to replicate the experiment. It is not the situation with this report where the methodology is written in a complicated way with a mixture of discussion and reported to other works. This does not make clear the separation of progress achieved by authors.

The authors weren’t able to separate the results of the discussion. Moreover, it seems that the teams don’t understand well the necessity to separate the two chapters. Always new research brings a small improvement compared with the last, and the readers need to know which is this release. For this, we must cite the results of similar works and compare our results with this, and this is the main purpose of the Discussion chapter.

In this way, a lot of results couldn’t have a better interpretation. For example, the box and whisker charts, Figure 10, 13, need a statistical interpretation regarding the location of the data in the box quartiles

Since I first read the manuscript there was something obvious that I passed, it is like an elephant hidden in the room: Is this paper fitting the journal's aims? Maybe its place is in a journal dedicated to Remote Sensing!

I concluded that the authors need a longer time to organize their results in a more readable way.

minor checks

Author Response

The authors appreciate the time taken by Reviewer 2 to assist in the improvement of our manuscript. As suggested by the reviewer we have made the following improvements:

• Reviewed the methodology to ensure the steps taken in this project are  clear and can be replicated. This was assisted by additional external reviews of the manuscript.
• We have discussed the separation of the results and discussion with the editor and have decided to leave them combined as we believe this suits our manuscript and may introduce confusion with the reader if the results are not discussed when reported.
• We have updated Figures 1, 9, 10, and 13 including more description regarding the box and whisker charts.
• Additional minor edits to the manuscript to improve overall readability.
• The authors believe the manuscript is a fit for the journal and the special issue we have submitted it for.