Hyperspectral Anomaly Detection with Auto-Encoder and Independent Target
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper uses auto-encoder and independent target technique for hyperspectral anomaly detection. The organization of manuscript is clear. The proposed method is tested on five public datasets. There are some important problems to be considered.
1. For the equation (1), it consists of three components. It is better to add the constraint for each component. For example, what norm is used to constrain sparseness、
2. The window size is an important parameter for detection result. It is better to analyze the parameter influence on detection performance and the parameter setting principle.
3. From the experimental results, it seems like the performance of the proposed method on the Hyperion Scene, San Diego Airport Scene and Gulfport Scene does not get the best accuracy, and the whole performance is not competitive when considering both the accuracy and efficiency.
4. Besides, as this method is based on deep learning, so the HAD based on deep learning should be compared in the experiment.
Comments on the Quality of English LanguageThis paper uses auto-encoder and independent target technique for hyperspectral anomaly detection. The organization of manuscript is clear. The proposed method is tested on five public datasets. There are some important problems to be considered.
1. For the equation (1), it consists of three components. It is better to add the constraint for each component. For example, what norm is used to constrain sparseness、
2. The window size is an important parameter for detection result. It is better to analyze the parameter influence on detection performance and the parameter setting principle.
3. From the experimental results, it seems like the performance of the proposed method on the Hyperion Scene, San Diego Airport Scene and Gulfport Scene does not get the best accuracy, and the whole performance is not competitive when considering both the accuracy and efficiency.
4. Besides, as this method is based on deep learning, so the HAD based on deep learning should be compared in the experiment.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsFig 1: The middle row and the diagram on the far right，
seems to be of little use from the perspective of the process。
Line 237: What are the meanings or possible meanings of the Six versions of distance measurement for multi-scale anomaly detection . why does the author requires the use of the six.
Line 253 and Line 256: The serial numbers of both formulas are 11, so we need to rearrange the serial numbers
Line 249 and Line256 : The formulas for the two distance measurement are the same, So The formulas need to be corrected.
Line 258: After this, in addition to mathematical formulas, some qualitative explanations for each formula may be provided, so that it be more beneficial for reading.
Line 316 AUC is a very important indicator, the formula for different indicators of AUC should be given。
Table 1 Will the values of the parameter such as m,p and h affect the final result? What are the reason for setting them. Whether is based on the final result selected in the experiment or a reference to a certain literature。
The Result of these experimentations and analysis are a bit like keeping a journal, you can find a main line to analyze. Different methods perform differently in different scenes and indicators, and different detectors of AE-IT also perform differently in different scenes and indicators. Is there any regularity in this conclusion that can help select methods and detectors
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe paper presents an unsupervised methodd using Low Rank and Sparse Matrix Decomposition (LRaSMD) for anomaly detection in hyperspectral data. The main novelty of this work is the use of a neural network with an Autoencoder network that improves background reconstruction. The work shows results using different distance measurements and compares with other methods. In my opinion this work has two main issues:
- It is only a small incremental improvement over other LRaSMD approaches.
- The presentation of the results is really confusing. Even the choice of the metrics to compare results is confusing (although adapted from another reference).
I think this article needs work to improve these aspects before being considered for publication. Among other suggestions for a revised version:
- Introduction does not consider some other methods and references seem to include very few different authors
- There are no details on the AE network use. The choice in the number "m" also lacks explanations (only reference with method)
- The explanation on how the AE-IT algorithm works could be made more clear
- The presentation of the results is not clear. Actually explanations of the meaning / importance of the different ROC curves is not explained (forcing the reader to go to reference 20). It is not defined what is AUC_(D,F). It is not explained why so many different metrics are needed.
- It is not explained why to use 6 different measures. Was this considered as an ablation study? Is there one that shall be used? I would suggest to include the 6 metrics after the comparison with the other methods.
- In general the results of the AD do not seem good if they are compared with other different approaches (pure DN approaches, even using AE).
- It is not possible to read the legend, axis numbers or axis titles of the ROC curves. The meaning of Tau, might be easy to deduce but it is not explained in the text
- The method shows high variability between the datasets, the authors provide an explanation, but there is little evidence supporting that explanation. Perhaps some more studies could show more evidence.
Comments on the Quality of English Language
In general small corrections are needed in the text. Examples are use of word "date" instead of "data" in the first sentence of the abstract. The repetition of "AEIT-based" in the conclussions or use of expressions like "excavate". Sometimes it is also a bit difficult to follow the text, some re-writing could help to make the text more easy to follow.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThanks for the revised MS, I think the revised contents have fully responded to my comments.
