Marine Radar Oil Spill Extraction Based on Texture Features and BP Neural Network

Round 1

Reviewer 1 Report

The objective of the paper proposed by the authors is to introduce an approach based of analysis of texture features and neural networks to extract oil spill from images measured by marine radars in the X band. 

The previous methods for a similar objective are widely detailed in the introduction. The reviewer is not a specialist in this domain but the overview seems to be complete. Following this introduction, the proposed method is well explained. First the authors introduce the classical problems that appear in the radar images mainly due to co-channel interferences. The choice of the Otsu threshold after a first convolution process that is a dynamic one seems to be pertinent. The images shown in this article show clearly the efficiency of the process. Then after, the process of cleaning the images, the authors introduce the image texture features determination. For this objective, the Grey Level Co-occurrence Matrix (GLCM) approach is retained. Such a method has already been used by authors as Lie and al. In this present work, the authors use 14 texture given by the table deduced from the GLCM method. Then a PCA method is an interesting way to compress the information with keeping only the main influence components. After that, a Neural Network is constructed. In this part, some information are not present and merit to be completed. These information are for example: the number of layers, the activation chosen for each layer, the loss function chosen in that case (for classification purpose). These information are given in the particular example in the following but as it is a classification problem some of them are imposed.

Some interesting results are then given. The process is presented on such examples. This conduce to the expression of the 2 main PCA components versus the 8 chosen features. The results are very good and give expected results. 

Finally a discussion is made, one point of this discussion concern the choice of the classifier. The K-NN approach seems to be the best and the tree algorithm the worst. Do the authors try the random forest that is often very efficient for such kind of problem ?  

This conclusion has to be moderated, as said by the authors, when using an adaptative threshold. 

As a conclusion, this paper is very interesting and clearly written. The results are good and the method  well described. I recommend this aper for publication. Some minor points quoted above should be improved.  Another point to modify concern the legend of the figure that are written in very small characters and difficult to read on a printed version f the paper. The authors have to modify this point before publication. 

 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

It is important to have wide area oil spill monitoring (detection) system using radar and optical to make a plan to remove oil spill.  In this content, your paper is very interesting and important. However, I have several comments and question as follow:

 

1.      To understand the data quality, it is very important to explain X-band marine radar system. So, can you add any information of X-band marine radar system? The reason why I ask is that azimuth distance effect should be considered (like as incident angle and distance from radar system, etc.)

2.      How about weather condition during X-band observation? X-band data were strongly affected by weather condition (sea wind as well as rainy condition).

3.      In table 2, why is principle component 1 no good enough when we compared with other components?

4.      How to select sample / training point?

5.      Figure 9 and 10, there are changed K-NN and decision tree (b) and (c). Is it correct?

6.      There are no enough explanation of difference method extraction results when we compare K-NN, decision tree and BP. Can you add your more interpretation about those and why it should be happened?

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

No futher comment