Spatial–Spectral Joint Hyperspectral Anomaly Detection Based on a Two-Branch 3D Convolutional Autoencoder and Spatial Filtering
, Siwei Zhao 1, Dandan Li 1,2, Boyu Pang 1,2, Xiaoying Lian 1,2 and Yinnian Liu 1,2,*Round 1
Reviewer 1 Report
The main contribution of this work is the development and performance evaluation of a novel algorithm for hyperspectral imaging anomaly detection in an unsupervised manner. The network is based by an autoencoder with two branches that uses spatial-spectral joint and spectral features. Model is tested an compared with six state of the art methods in airborne and satellite borne hyperspectral datasets, showing improvments in the detection of anomalities and background superssion.
Introduction includes a logical revision of anomaly detection methods and their principal characteristics.
Contributions are clearly stablished and network architecture is described in detail.
Characteristics of datasets are well described, although ABU dataset (line 337) spectral interval is not specified.
Results show the improvement of the proposed method compared with selected methods based on three evaluation criteria.
Results are limited to the experimental settings, for example selected datasets and evaluation criteria. It could be discussed diferences between results of the two kinds of datasets (airborne and satellite borne).
Author Response
Dear Reviewer:
Thank you for taking the time to review our manuscript and providing valuable comments and suggestions. We have made revisions to the manuscript based on your suggestions and have responded to your comments point by point. Please see the attachment for the detailed responses.
Best regards,
Author Response File: 
Author Response.pdf
Reviewer 2 Report
See the attached file
Comments for author File: Comments.pdf
Author Response
Dear Reviewer:
Thank you for taking the time to review our manuscript and providing valuable comments and suggestions. We have made revisions to the manuscript based on your suggestions and have responded to your comments point by point. Please see the attachment for the detailed responses.
Best regards,
Author Response File: Author Response.pdf
Reviewer 3 Report
Critical response on manuscript entitled “Spatial-Spectral Join Hyperspectral Anomaly Detection Based on Two-Branch 3D Convolutional Autoencoder and Spatial Filtering” by Shuai Lu and others submitted to
The paper under consideration is the work of merit. It is dedicated to automatization of hyperspectral image processing and object detection using original two-branch autoencoder and spatial filtering. However, the description of the applied method, research technology and software resources lacks details. Also, there is no availability of source codes and study datasets which is necessary for reproductivity of the results and could be of interest for international readers.
Firstly, the meaning of “anomaly” in case of satellite image processing and brief literature outline should be made (lines 41-44). In which field described technology is applicable (abstract and Intro don’t have it)? Because anomalies are always subjected to environment, in natural or urbanized landscapes it could be completely different. Lines 46-60 are related to spectral anomalies mostly, and not applicable to natural anomalies like lineaments and circular structures of purely relief anomalies (like mesa, caldera etc). Research background and works of predecessors are not clarified.
It is not much convincing, why researches has choose ANN to any other methods? (lines 99-150). Used software also not specified. Stages of the proposed method (Fig. 1) should be enumerated and properly referenced in the text. Some legend describing meaning of colored blocks is needed. Algorithms description is too wordy. Selection of model parts and their configuration is unknown. Why this configuration is the most advantageous relatively to the others?
Details of Fig. 4 are hard to read. There entire section 2.2.2 needs explanation. Why did authors choose 3D-CAE. Is it original invention? Were any similar developments ever used? In which fields?
Section 2.2.3 obviously needs description of training/datasets and model validation. Did authors used anomalies outlined/labelled by humans? And, important, description of the limits of model application is unknown (could model be exported to other areas?).
Lines 327-360. Datasets with very limited environments are mentioned. Moreover, achieved results (fig. 5-8) could be reached with simplest Outsu threshold method or unsupervised clustering algorithms (K-means, for instance).
Conclusions. According to general impression I would recommend major revisions, but manuscript could be declined as well.
Author Response
Dear Reviewer:
Thank you for taking the time to review our manuscript and providing valuable comments and suggestions. We have made revisions to the manuscript based on your suggestions and have responded to your comments point by point. Please see the attachment for the detailed responses.
Best regards,
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Critical response on revised manuscript entitled “Spatial-Spectral Join Hyperspectral Anomaly Detection Based on Two-Branch 3D Convolutional Autoencoder and Spatial Filtering” by Shuai Lu and others
In the revised manuscript under consideration most of my past recommendations were fulfilled. The text of the manuscript was mainly improved in a way, which made it more comprehensive and interesting for the international reader.
Authors mainly explained emerged questions. The English of manuscript was considerably improved. Comprehensive explanation was given to methodology and data processing. And, fortunately, used software libraries and source codes were added into data availability section.
The manuscript as it presented by now doesn’t require any serious amendment and could be published in a present form.
