Hyperspectral Anomaly Detection Based on Multi-Feature Joint Trilateral Filtering and Cooperative Representation

Round 1

Reviewer 1 Report

Hyperspectral anomaly detection based on multi-feature joint trilateral filtering and cooperative representation

This paper proposed a novel algorithm which detects anomaly in hyperspectral images. The algorithm first applied an improved trilateral filtering to remove outliers from an image, taken into account its spatial features, and to have its background filled. The results were then represented collaboratively to detect the actual anomaly. Anomaly extraction by subtraction was by no means novel. However, its contribution was utilizing spatial characteristics of the underlying features via filtering. Accordingly, the reported performance of the proposed background suppression was impressive. Its objectives and solution were clearly presented. The algorithm design and evaluations were technically sound. It was benchmarked against seven state-of-the-art algorithms. The resultant evaluation metrics indicated that it marginally improved the reconstruction performance of involved hyperspectral remote sensory images. After the following issues and concerns are carefully considered and addressed, it is believed that the paper is worth considered for publication in this Applied Science.    

Major Points

1. Section 2.1. Data, please provide that definition of \anomalies considered in this paper, especially how did they differ to other “normal” objects. What are the spectral (or spatial) characteristics that differentiated, for instance, road tracks from the vehicles on them? (Was it size, shape, etc.?).

2. Following the previous point, if it was the size, doesn’t it mean that the window sizes (Page 4, Line 121) and standard deviation between adjacent pixels, their intensities, and gradients (sd. in Eq. 9, Eq. 11 and Eq. 13) are determining parameters. In that case, please provide a guideline on how to specify these (and other user) parameters.

3. Figure 4, it is not clear in the section 2.3.2 whether the selection of a reconstructed window was user defined or automatic. Please clarify and provide more explanation.

4. Following points number 2 and 3 and details given in Section 3.1., (1) to 3.3 (1) Parameter Setting, please provide another experiment to confirm the invariability of the results against inter-and intra-observer choices of the empirical parameters (within given ranges) in practice.

5. Since the AUC results reported in Tables 3 and 4 seems to be only marginal improvement over the existing methods. Please discuss other benefits (probably speed or robustness, etc.) of the proposed algorithm that outperformed them at much larger degree and thus justify its adoption over the others. In that case, additional experiments to collaborate such statement are also needed.

6. Please provide visual examples (i.e., original hyperspectral images and detected anomalies) of the false alarms found by other methods, but correctly detected by the proposed one. Similarly, please also provide some visual examples that the proposed method missed. In addition, please discuss these cases in detail before Section 4.

Minor Points

1. Since the audience of Appl. Sci. is Scientist in various fields, abbreviations of specific terms (e.g., RX algorithm, etc.) should be given. For instance, an abbreviation table at the beginning or as an appendix is welcome.

1. Academic writing style in this paper needs to be enhanced. Additionally, minor grammatical errors, redundant words and phrases, and awkward choices of vocabulary and tortured phrases were found. Please revise, preferably by a native English speaker.

2. For instance, the phrase “background purification” seems like a tortured phrase. Please provide its description and maybe choose another more appropriate term (used by other prominent papers)

Author Response

The detailed responses are all in the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

In general the paper could be interesting, but many parts are unclear and not defined. There is not enough explanations, in many parts, about what is that the authors are really talking about. For example the talk about "background dictionary" even in the abstract, but they never define it. Or they talk about "detection properties" but is not clear what they mean. 

Additionally, they talk about multiple datasets without giving any reference, this should be addressed as it is fundamental (see section 2.1). The don't explain concept as "size of the inner window" (what is an inner window?) (in line 121) and many others during the paper.

One major point is that is not explained how the datasets have been splitted in train/test and how this has been used by the authors when giving results. This needs to be addressed.

Additionally I found that the conclusions are a bit too optimistic, and are not completely supported by the results. They have (for example in Table 2) a results of 0.9998 that should be compared with the next best value of 0.9965. The difference is surely not statistically significant, and thus claiming that their method is better is a bit of a stretch. A statistical analysis is in order here (for example by splitting the dataset in multiple ways).

All in all I think the paper could be interesting but there are much work to be done to bring it to a form that could be publishable.

English should be improved. There are some typos in the document that needs to be addressed (like author names that are wrong, see for example line 45: "Matteo li").  

Author Response

The detailed responses are all in the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have satisfactorily addressed my concerns. I have no further academic suggestion on the manuscript.

Language quality is good and usual editorial proofreading process is needed.