A Deeply Supervised Attentive High-Resolution Network for Change Detection in Remote Sensing Images

Round 1

Reviewer 1 Report

This manuscript proposed a deeply supervised attentive high-resolution network (DSAHRNet) for high-resolution remote sensing images change detection. The theoretical description of the paper is very detailed, and the ideal is interesting. In the experimental results section, the proposed method outperforms other compared methods. However, some questions that should be concerned

(1)    The motivation and contribution of the manuscript need to be further clearly emphasized.

(2)    Some STOA methods regarding change detection should be compared in your experiments.

(3)    Some more methods regarding remote sensing should be investigated in your introduction, e.g., Semi-Supervised Locality Preserving Dense Graph Neural Network With ARMA Filters and Context-Aware Learning, Graph Sample and Aggregate-Attention Network, Unsupervised Self-correlated Learning Smoothy Enhanced Locality Preserving Graph Convolution Embedding Clustering, Self-supervised Locality Preserving Low-pass Graph Convolutional Embedding, AF2GNN: Graph Convolution with Adaptive Filters and Aggregators, Deep hybrid: multi-graph neural network collaboration for hyperspectral image classification .

(4)    The authors should experiment the adaptability of the proposed algorithm to different numbers of labeled images. In addition, are the comparison methods tested in the same environment? Further clarification is required. From the existing comparison results, the method proposed by the authors has poor adaptability to small label data, which directly limits the practicability of the algorithm.

(5)    The authors should use experiments to illustrate how to choose the hyperparameters of the algorithm.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The submitted paper presents a work, where a Convolutional Neural Network is applied on the Change Detection of Remote Sensing Images.

The layout of the presentation of the research is classic: a readable and well-structured introduction (1) is followed by a state-of-the-art section (2)
with a sound presentation of the own work done (3). The algorithm then is applied on a set of well-established datasets with a presentation of the results (4),
followed by a discussion of the latter (5) and a brief conclusions section (6).

Clearly, the point why this work should be published is, that it shows tiny improvements of other algorithms that have been applied on the datasets before.
If this would not be the case, it would be just "yet-another-neural-network-paper".
When going through the paper one finds a very well elaborated and presented methodology with adequate experiments.

Still, some points could be made clearer. Where a concrete advice can be given, it is directly noted, and formulated as a question if this was not possible:
- p.1, l. 31: as MRF are only the mathematical model the "classifier" to name in context of [20] would rather be "Spatial Domain Analysis" and not "MRF",
- Section 2: there is no 2.2, so there should also be no 2.1. Furthermore, as the authors distinguish between two categories of CD networks, they should order theirs into one of the two.
- Section 3.1: The reader is pretty much left alone with the "multi-level features": in order to allow a publication, it has to be made clearer what hey actually are.
- Section 3.1: Please add a reference to the loss-functions, which are well-described later in 3.3. At the moment, when they are introduced, the reader is unaware of their definition.
- Section 3.3: Maybe a reference back to the explanation of the algorithm described in 3.1 would be adequate.
- Section 3.3: The definition for the distance measures d_i and the classifier M is somewhat too vague. This has to be specified clearer. Again, as for the the note about the features this has to be considered crucial for publication.
- Sections 4.6 and 4.7: In my eyes these discussions are pretty lengthy with not very meaningful tables and figures, that could be overdone in my eyes, being more concrete.
- The References need to be overdone for a consistent presentation.     

Minor errors can be corrected as well:
- p.3, l.140: learn -> learning
- p.4, l.146: represents -> is represented
- p.4, l.164 (twice): respresents -> represent
- p.5, l.187: back-propagating -> are back-propagated
- p.9, l.280: ~ -> -
- p.10, equation (15) (twice): Presicion -> Precision
- p.11, l.362: its -> in a
- p.11, l.362: such phenomenon -> such a phenomenon
- p.12, l.426: been -> be
- p.12, l.431: are -> is
- p.16, l.460: achieves -> achieve
- p.16, l.462: form -> from
- p.16, l.483f (three times): base -> based
- p.17: l.499: constructed -> is constructed

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper proposed a supervised attention high-resolution network for high-resolution remote sensing images change detection. In general, the technique is interesting and the paper is well-written. My comments are as follows:

1.     If the spectral and spatial attention module is not added to the model, how about the visualization effect of attention map (compared to Fig.5(b) and (c)?

2.     How about the stability of the model when repeated experiments? (Variance value)

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have made great efforts to solve the problems I raised, but some problems need further elaboration, mainly as follows

1.The authors should use experiments to illustrate how to choose the hyperparameters of the algorithm；

2.Since the method proposed by the authors has poor adaptability to small label data, compared with the SOAT semi supervised method, what is the competitiveness of the method proposed in this paper? This is critical. The authors must prove it by experiment；

 

3.In addition, the method proposed by the authors is an improvement based on CNN method. There are already related similar methods. What is the innovation degree of the proposed method? Now the innovation degree is not enough to be published on RS. The authors must pay attention to this point；

4.It is recommended that professional organizations polish the language；

5.The authors felt a little perfunctory about their comments.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 1 Report

No more comments. The paper can be accepted as presented form.