Field Patch Extraction Based on High-Resolution Imaging and U2-Net++ Convolutional Neural Networks
Round 1
Reviewer 1 Report
This manuscript proposed a method of using the U2-Net++ model in combination with remote sensing data for accurate field boundary extraction. The new built U2-Net++ model based on the RSU module was effectiveness of field extraction. We recommend receiving the article after minor revision
Suggestions:
1. This manuscript selects three different types of cultivated land (drylands, paddy fields, and terraced fields), however, they all seem to have a fairly regular boundary. In practice, many cultivated land boundaries are irregular and may be heavily obscured. Does the paper consider the extraction of irregular boundary based on the new method and the extraction accuracy?
2. The method part does not explain how the training set, validation set and test dataset are allocated to the samples. Does the test dataset used for precision verification participate in the training?
Moderate editing of English language required
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Authors proposed a U2-Net++ model for accurate extraction of farmland boundaries. I think the ms is generally well-written. I think the ms can be considered for publication after a minor revision.
Majors:
· Figure 1: replace background map that displays a part of the disputed area. This may upset international readers.
· Improve quality (DPI) of map figures as they look fuzzy and have many broken texts.
· Given that the accuracy of U2-Net++, U2-Net and Deeplab V3+ are not much difference. It would be great to assess the computation cost of those models as it would be vital for an operational mapping task.
· Since this is a method paper, I encourage the authors to make their codes / data publicly available.
Minors:
· Line 53: remove “such as…”.
· Line 205-206: typo “U-Net++”, Figure ???
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The Paper discusses the implementation of various deep-learning segmentation algorithms for the boundary detection of the remote sensing Images. The data set used in the study is downloaded and labelled by the authors.
1. The data set description is not appropriate; this does not describe appropriately about the dataset. In section 2.2, the source and resolution are mentioned, like GZ-7, Google Earth etc. but then how these different sources are used. are they jointly used to label etc... is not mentioned
2. The use of ArcGIS for the label creation is done, which also needs some clarity as calculating the centroid and then clipping images of size 512X512, makes no sense; why not cut the grid?
3. The accuracy of the created labels needs to be established. Are the software-generated labels (boundaries) in line with the ground truth?
4. Are you sharing the dataset? if some researcher needs to work on this data is it sharable? how?
5. None of the figures are cited(explained with references), if at all, the written " as shown in the figure below.." This is not the right way of citing the tables and figures in the paper.
6. More details of the data set is mentioned in the result section too, which can be grouped with the dataset section.
The language of the paper needs correction and more in the flow of the paper. The Introduction section needs to be subdivided for a better-structured introduction section.
The statement "The edge detection method uses edge information within the remote sensing image to extract the field boundary by analysing the relationships between edges." is confusing and hardly makes any sense .
Author Response
Please see the attachment
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
Most of the Suggetions are addressed in the response sheet.
Look forward to the followup commitment of providing the labelled data for the publically available data.
