Potential of Convolutional Neural Networks for Forest Mapping Using Sentinel-1 Interferometric Short Time Series

Round 1

Reviewer 1 Report

Please see the attached document

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript applied U-Net-like CNN for mapping map rainforest areas threatened by landscape changes and degradation using sentinel-1 InSAR short time-series under varying inputs. The study of this manuscript is interesting, which assesses the potential for mapping the land cover over the dense-cloud cover area. The paper was well written and neatly organized. I have some minor comments for authors to consider.

I am confused by the terms “short-time-series”, what’s the point of “short-time-series” in this paper? any particular meaning in this case study? Does it mean a short revisit time of Sentinel-1? Or you are able to map the landcover by such a short revisit time S-1 InSAR data?

Figure 2. “Our goal is to reduce the processing time required for classifying S-1 InSAR images by skipping the texture estimation and exponential model fitting steps.” Does this study need to prove this?

Apparently, local hierarchical features are learned by the stacked convolutional layers in CNN models, which reduces uncertainties in feature selection (such as the feature engineering in RF or Adaboost) and improves automation in semantic labeling.

Line 321, details the 18 texture feature and their equations.

Line 310-337, this part is mainly about the experiment design, should be moved to the method section.

Line 354, since the random forest was applied, you need to clarify the parameters in RF modeling.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper is an interesting look at the use of CNN and SAR data to map forests, a follow on to previous work in the same area of Brazil which used Random Forests with SAR backscatter. Overall, the paper is interesting, and of a high quality. However, I do have some concerns, generally with increasing the level of specific detail given, and with interpretation of your results, which are listed below. Main points: The Accuracy rates/F1-scores for Case III and Case X are sufficiently close that I think you should test for statistical significance – i.e. For Case X you get F-score of 92.85% vs 92.5% for Case III. Is this a statistically significant improvement? You should test it. I note that Case III beats all Case Iv – IX. Furthermore, given how close Case III and Case X are in terms of accuracy/f-score, I suspect that the use of textural features and use of CNN (ie Case II but with CNN) might well beat it. Did you test this case? You are claiming that switching your method from use of textural features to coherance stacks improves accuracy. I am not sure that currently your results bear this out. Table 4 I’m curious about the (relatively) poor accuracy figure in Case X for water (95.58%), I think it should be 99.58… Abstract You should supply numbers – ie give scores for case III and case X in the Abstract.(ln12-14) . I have an additional few small points/additions which I list below. Minor points: Ln135 “Figure 1, The…” Comma instead of full stop Just to be clear, you didn’t use any ascending images? Ln176 “the latter is a reasonable assumption…” Confusingly worded – what is the former assumption? “This is a reasonable….” Would be better Fig 2 could be more detailed. SAR processing and InSAR processing boxes are not very informative, and could be subdivided into substeps. I’d be interested in a few practical details of what software/processing steps is used, to be added to Section 2. SNAP? Python? This comment applies also to section 3.1 and the description of the U-Net. Section2.3 Ln198-204 I found this paragraph unclear – what is the resolution of the FROM-GLC map – 10m or 30m? “The result is a stable classification, as less than 1% overall accuracy is lost when less than 40% of the training set is used.” What is this sentence referring to – what training set? Ln230-232” The proposed patch-based augmentation includes both horizontal and vertical flipping, as well as rotations with mirrored padding on the borders.” Citations for these techniques would be helpful Ln237 “as can be seen in Figure 5.” Section 4 I do think a section called Experimental setup (section 4.1) would be better found in materials and Methods than Results. You shouldn’t be explaining how you carried out the experiment in the Results. Ln399 Worth giving abbreviation for NDVI Ln405 “false alarms” false results would be better Ln408 “This is due to the fact that the random forests classifier solely relies on pixel-wise estimations, while the CNN is able to consider the neighboring spatial content of the data.” Not true in case of use of textural features, unless I’m misunderstanding your use of pixel-wise estimations.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Dear Authors,

I have reviewed the paper titled: “Potential of Convolutional Neural Networks for Forest Mapping using Sentinel-1 Interferometric Short-Time-Series". In my opinion, the aims of the paper are germane with “Remote sensing” journal topic, however, in the present form, the paper fits only in part with the international scientific and journal standards. The paper is written with a good English level. The contribution of this paper to the scientific knowledge is acceptable but some flaws are present in the text. I understand the difficult work done, but as a reviewer it is my duty to highlight the gaps in order to improve the research approach and its presentation to the international scientific community. Please I suggest revising the paper following the suggestions and comments reported below:

• The one used does not seem the template of the journal, in my opinion there are some mistakes, please check and correct;
• I think that literature references should be removed from abstract and consequently the references numbers in the text must be re arranged;
• Missing scale bar in figure 1, and in all other figures, please add;
• Conclusion section has to be further focused on the main results according to the main goal.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

No further comments.