Differentially Deep Subspace Representation for Unsupervised Change Detection of SAR Images
Round 1
Reviewer 1 Report
General comments
An unsupervised technique for bi-temporal SAR images change detection is proposed in this paper.
An interesting approach is proposed; however, this paper seems lacking in two respects.
First, it seems failed to provide sufficient details on how the computation was realized. This concern is crucial in ensuring the reproducibility of the manuscript methodology.
Second, the inherent effectiveness of the proposed performance evaluation appears questionable, thus it needs significant improvements, as discussed below.
Specifically, the experimental dataset is limited, and the evaluation of the performance variability with respect to algorithm parameters (e.g. the number of hidden neurons) appears not fully representative, insofar as the analyzed scenarios are extremely limited. Moreover, it is not clear neither the information content inherent to the selected reference change maps nor how the training specifically affects the performance.
Therefore, I would abstain from recommending this manuscript for publication in the present form; however, I encourage authors to improve their manuscript according the reported comments.
Detailed comments
Generally it is not suitable to report the reference numbers in the abstract; Authors repeatedly use the “speckle noise” term throughout the manuscript. I would highlight that, strictly speaking, the speckle in SAR images is not a noise (please refer to the classical textbooks for the SAR fundamentals); The adopted terminology is sometimes obscure (see “Unpredictable difficulties,” line2; “fully of strong speckle noise,” line31).
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors developed a change detection method for SAR images based on the combination of AE-like network and self-expressive layer. The results are fine in the case. The detailed comments are following.
1. p.2, Introduction
The authors mentioned that the difficulty of change detection for SAR images is due to speckles. But the no statements will appear about the speckles in the methodology. Why can the proposed method deal with the problem? It is better to describe the relationships between the difficulty and proposed method.
2. p.3, line 101
What is the reason that they are not well adapted to the remote sensing change detection tasks?
3. p.3, line 111
What is the advantage and disadvantage of the self-expression layer compared with SFA base linear transformation? It is required to state clearly the reason of introducing the self-expression layer instead of SFA.
4. p.8, Table 1
In the case of patch-wise strategy, how did you identify the TP/FN/FP/TN? The definition of the elements is necessary.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
A deep learning architecture for detection of pixels in multi-temporal SAR images which are changed during time is proposed. Architecture combines autoencoder network together with linear orthogonal transformation in the latent space which is learned during training of prosed architecture. Distance between the orthogonal latent representation of image patches is clusterized in two clusters. According to obtained clusters a decision about pixel change is made.
Although the idea conceptually is well presented there are lack of information about some important steps in method explanation and in experiment evaluation. In the following I give few remarks on which authors
must pay attention.
First, motivation behind orthogonality of linear transform in the latent space is not well explained. Why orthogonality is such a desirable condition and how this is related to problem considered? It would be of great importance to clearly present this reasoning since according to authors current presentation this is among the main contributions of the work.
Second, explanation of terms in loss function should be better, particular normalization term used notation for variance which is not previously ntroduced and it is not clearly explained why this formed of normalization term is proposed. Regarding the experiment evaluation my main concern is related to determination which cluster of points is denoted as unchained and which as chained pixels ?
Also I would like to see what prediction algorithm gives for those pixels for which you don’t have groundtruth (the ones that are denoted with black colour in groundtruth map)?
Another concern is related to independent search for the number of hidden layers and the number of hidden neurons. The joint search for both parameters in a grid like scenario (the same that you used for one parameter) can give much more useful information, primarily about importance of each parameter and also about their might possible correlation.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have somehow addressed the points raised. The new version of the manuscript deserves to be published in the present form.
Reviewer 3 Report
The authors adopted most of the suggestions proposed by reviewers in the first round and significantly improved the manuscript. One last requirement from my side would be to increase the font size in the graphs in figure 5.
