Feature Selection for SAR Target Discrimination and Efficient Two-Stage Detection Method

Round 1

Reviewer 1 Report

This paper proposes a two-stage detection framework including feature selection for SAR target discrimination. The majority of detected clusters of false detection can be mitigated successfully in the experiments, which is nice.

Before publication, there are still some important work to do. Some of my concerns are presented below.

1. The two-stage detection method is proposed in this work. In the experiment part, it is strongly suggested that a few state-of-the-art detection approaches should be included as performance comparison.

2. The terminology in the manuscript is a bit confusing. The target discrimination is a step of the whole target detection procedure. Fig. 1 shows the whole detection flowchart including the target discrimination scheme. Please properly modify the caption of Fig. 1.  The caption of Fig. 2 is incorrect. Please check.

3. Feature selection and transform are the core of the detection method. In order to verify the effectiveness, it is better to provide a comparison between the discrimination performance using selected features and the discrimination performance using a single feature.

4. Please provide the ground truths of the targets in the tested images if possible.

5. According to Table 1, the probability of clutter exclusion is 0.908 after the coarse discrimination. However, a large mass of clutter still remains in the result after coarse discrimination, which seems to be far more than 0.1 of the original clutter. Please explain in detail.

Author Response

We are grateful for your sincere review. Your comments have greatly helped the development of the article.

Please see the attachment. 

Author Response File: Author Response.docx

Reviewer 2 Report

Feature Selection for SAR Target Discrimination and Efficient Two-stage Detection Method

General description

A two-stage detection framework is suggested to ensure efficient and high detection performance in Terra SAR-X (TSX) images based on previously studied features. The proposed method consists of two stages. The first stage uses simple features to eliminate misdetections. Next, the discrimination performance for the target and clutter of each feature is evaluated and features suitable for the image are selected. Karhunen-Loève (KL) transform is applied to reduce the redundancy of the selected features and maximize discrimination performance. The proposed method allows the majority of detected clusters of false detections to be excluded, and the target of interest to be distinguished.

Remarks

Row 152-153. The sentence has to be corrected: “In this step, the targets of interest and the clutter are discriminated against based on the features introduced in [13] and [14].” Replace “against” with “again” or remove again.

It is recommended the flowchart in Fig. 1 to be detailed with structural equations (4) – (9).

It is recommended English to be carefully checked in the entire article.

Author Response

We are grateful for your sincere review. Your comments have greatly helped the development of the article.

Please see the attachment. 

Author Response File: Author Response.docx

Reviewer 3 Report

In this manuscript, the authors propose a method to obtain good performance for discrimination between target and clutter and describe quantitative evaluation of the performance, which merits publication in Remote Sensing. For the benefit of the reader, however, a few points need clarifying and certain statements require further justification. These are given below.

1.      3.2 Preprocessing and Coarse Discrimination Step (Lines 210-214): Several parameters are chosen, for example, a false alarm probability of , the neighborhood range ,  and MinPts of 3. Please explain the reasons for choosing these numbers.

2.      Line 149-150 and Table 2: There are too many abbreviations. Please explain them. You could list all of abbreviations at the end or first.

3.      4. Discusson (290-295): It is stated that some adjacent targets were clustered into one object. Please discuss this point in the text giving an example. 

4.      Table 1: Please explain the reason why all of “Recall” equal to be 1.0000. 

5.      Line 253: Eq (6) could be Eq. (8).

Author Response

We are grateful for your sincere review. Your comments have greatly helped the development of the article.

Please see the attachment. 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

All my questions have been nicely answered. I would like to recommend it for publication.