Deep Residual Autoencoder with Multiscaling for Semantic Segmentation of Land-Use Images

Round 1

Reviewer 1 Report

The author presented a new flexible autoencoder-based architecture of deep learning for robust semantic segmentation of remotely sensed land-use images. The manuscript is well written and provides good sound contributions to the field of remote sensing, particularly in the areas of land use classification. The related studies sections were informative and useful to understand the rest of the manuscript. After reading the manuscript, I only have one comment: The section’s numbers need to be in order for example line in 128, 3.2 should change to 2.2. It needs to be check in whole manuscript.

Author Response

Please see the attachment. 

Author Response File: Author Response.pdf

Reviewer 2 Report

This article presents a novel methodological approach with demonstrated efficacy for semantic segmentation of remotely sensed images. Hence, this article is indeed worthy of publishing for scientific community. However, there are some minor revisions that will be necessary, as such:

Albeit this article is well-written, the current state of the discussion on the proposed methodology may not be well understood by the broader community, particularly those who will mainly focus on its applications. In that regard, the author should try to highlight on the methodological workflow in more cohesive and understandable way in Section 3, should there be no limitation on page. To partially take care of the point 1, description of a detailed workflow of the algorithmic implementation on GitHub (or elsewhere) would be necessary and helpful. Discussion on “Semantic Segmentation” itself seems somewhat inadequate. A separate paragraph can be added in the Section 3, solely focusing on this topic in geospatial application contexts. This will certainly be helpful for readers/researchers that are new to this exciting area of research, particularly within GIScience community. Line 508 (i.e. “The present results demonstrate that extensive …) has grammatical issue.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Author presents work that proposes a Deep Residual Autoencoder ANN, used together with multiscaling for land-use satellite image segmentation.

The article is insightful and it advances the state of the art in segmentation, according to PA and JI measures.

I propose a few places where the article could be improved:

Provide an intro to each section.  It is not a good idea to start 2.1 right after 2.

Define all acronyms at first use.

You are assuming the reader is completely familiar with ANN and DL terminology.  I do not think that is the case.  Define each DL scheme before providing examples.

Also, provide a more detailed explanation on your proposal.  How is it implemented?

Fig. 3. does not contain subfigures.  Do not call its parts a and b as if they were subfigures.

Eq. (1) What is this equation trying to convey?  Can you remove the ellipsis and complete the expression?

Symbol f is being abused.  It was used with and without subindices in Eqs. (1) and (2).

Line 298.  Training Samples -> Training Set

3.4. not clear (at that point) why oversampling, nor how.

Eq. (6) Jn has not been defined.  Define all symbols before or right after first use.

Fig. 8.  Plot histograms vertically, unless you have a reason and it is clear and explained.

please, provide a statistical analysis as to assert that one method is better than another one.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf