Hyperspectral Sea Ice Image Classification Based on the Spectral-Spatial-Joint Feature with the PCA Network

Round 1

Reviewer 1 Report

This paper focuses on extracting sea ice by using an Deep Learning (DL) based approach that assimilate spectral and spatial information simultaneously. The topic of this paper is interesting and the methodology flow is well presented. However, there are several major concerns from the reviewer.

 

1. The motivation of this research is not clear demonstrated. In the introduction section, the author first stated that there were some traditional supervised algorithms developed for classifying sea ice but the spatial/morphological information was not utilized. In fact, the traditional method such as SVM is able to take advantage of spatial information as long as we extract the spatial information and then give it to the classifier as the input. This is also what the author have done to the PCANet. There are quite a lot similar studies several years ago. For example:
2. Bruzzone and L. Carlin, "A Multilevel Context-Based System for Classification of Very High Spatial Resolution Images," in IEEE Transactions on Geoscience and Remote Sensing, vol. 44, no. 9, pp. 2587-2600, Sept. 2006, doi: 10.1109/TGRS.2006.875360.

 

10. Zhang and Y. Zhang, "Airport Detection and Aircraft Recognition Based on Two-Layer Saliency Model in High Spatial Resolution Remote-Sensing Images," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 10, no. 4, pp. 1511-1524, April 2017, doi: 10.1109/JSTARS.2016.2620900.

 

Therefore, the improvements that the authors made in this study do not sound that novel.

 

In addition, the authors also said that some DL based approaches have been proved to have the ability of extracting sea ice with good accuracy.  One strength of the DL is its capacity to automatically learn the features that include both spectral and spatial information from training samples. To be honest, I do not see the necessity of manually add such spatial information again to a DL based frame. The results in Table 9 confirms my suspicion: even though the proposed method obtained the highest accuracy, the difference between each method is too close. Given the fact that the method is only tested over quite a few scenes, it is hard to conclude that there is any progress made by the proposed algorithm.

 

2. The results section doesnot provide a solid proof to the readers for showing the superiority of the algorithm. For example, only some numbers related to the accuracy in Table 5 were given. The zoom-in classification maps along with the table can give the readers a much better understanding of the pros and cons of each approach. Another thing that we need to notice is that the labels used for evaluating the classification results are individual dependent because they are determined by human interpretation. Considering the tiny differences between the proposed algorithm and other benchmark methods, the rank of accuracy is also likely to be associated with how the validating samples were selected.  

 

Last but not least, I think the most meaningful question about sea ice extraction is how accurate we can distinguish ice pixels from different complex environments by getting rid of the impact from thick clouds, thin clouds, cloud shadows and sun glints which are pretty common in the ocean images. This may be pretty challenging especially during the ice forming and decaying seasons. In this study, the authors only focused a few scenes under clear weather conditions which is far from the really practical use.        

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors propose a method to perform sea ice image classification. The presented methodology relies on a custom feature extraction approach, fusing different features (texture, spatial, and spectral) that are provided as input to a PCANet. A final SVM classifier is exploited to perform the final classification. 
The presented approach was tested on two different datasets and compared with other classification algorithms, demonstrating higher accuracy.

The state of the art is well-reviewed, the target problem is well described, and the proposed methodology and the performed studies are well detailed.

The described approach is interesting and features some novelty. To further improve the work, I suggest the authors to compare the proposed methodology directly to the state of the art for at least one of the used datasets. This would make their scientific contribution more clear.

Finally, some typos are present in the manuscript:

-) At line 253, there is a missing reference

-) I believe the title of sections 3.1.1 and 3.2.1 should be data description and not date description

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

 

The submitted work proposes a hyperspectral sea ice classification approach that is based on jointly using the spectral-spatial features. After extracting the spectral and spatial information of hyperspectral data, a deep learning network called as PCANet is used to obtain the let’s say deeper feature information and this leads to a higher classification accuracy as well as a higher level of training efficiency. Generally speaking, the topic is interesting and to-the-point, as spectral-spatial approaches have been increasingly used and discussed during the past few decades (and that is why I am a bit surprised that many of the most famous methods have not been addressed/cited in this work). The novelty level seems fairly adequate, however, it is a combination of some widely-used and well-known tools and concepts. The structure of the manuscript is fine, but the language needs to be improved. In the end, I suggest considering the submitted work for publication after addressing my major/minor comments,

 

 

Major

• In general, the comparison section is weak since the proposed method should be compared to some state-of-the-art works which are dealing with the same concern that is Spectral-Spatial classification of HIS.
• How to set the parameters in a more hands-off or automated/adaptive way? It seems that the parameter settings are not following a consistent manner as several parameters should be re-set and modified according to the data. In general, that sounds reasonable to slightly change the value of the parameters according to the data characteristics. However, one main concern would be still tuning these parameters in a manual way. Any advice or solution?
• One problem with HSI data with the high spatial resolution is that the spectral-based classification methods cannot deal well with such data. One simple solution would be using a smoothing step as a pre-processing in order to decrease the spatial resolution. It could be also helpful if one intends to segment the image in object-based scenarios. Have the authors yet tried using a smoothing step as the pre-processing?
• Object-based image analysis (OBIA) is also widely used as spectral-spatial image classification (or in general ‘image analysis’) tool. Here is a relevant work:

Zehtabian, H. Ghassemian, “An adaptive framework for spectral-spatial classification based on a combination of pixel-based and object-based scenarios”, Earth Sci Inform,10: 357, 2017.

 

I would suggest addressing OBIA in the paper. OBIA (or as mentioned in the first comment, some other spectral-spatial classification works) could be also added to the comparison section.

• There are several typos in the text. It's recommended the manuscripts could be sent for proofreading to ensure the use of tenses and grammar meet the publication requirement.

 

Minor:

• When preparing the REFERENCES section, please follow the instructions defined by the MDPI Remote Sensing journal. Proper formatting of the references is crucial. The template is not consistent and there are typos and other issues with the references. For example, the publication year (2020) has been repeated two times in Ref. [2]. Also please provide the readers with DOIs if possible.
• It would be helpful if you add references to ALL the methods mentioned in the comparison tables.
• Please provide the potential readers with the full names of abbreviated terms when they are mentioned in the text for the very first time. One example is PCA in line 16 in Abstract.
• Line 19 in the Abstract: Please be consistent with the tense of the verbs in a sentence, as they should follow the same tense. Then please change ‘extracted to ‘extract’.
• Page 2, line 61: The Gabor filter IS based on the …. (is was missing)
• Page 2, Line 74: That is not always true.
• Page 3, the first paragraph of Section 2 (Proposed Method) is somehow a repeat of the first paragraph on the same page. I would suggest avoiding any redundancy.
• I also suggest using the abbreviated form of the term ‘gray-level co-occurrence matrix’ (i.e. GLCM) throughout the paper. (in Figure 1 it was stated as GLCM, but the full term has been used almost anywhere else in the text).
• Page 4, in Figure 1, I would modify the title of the last step in the flowchart. It should be Classification rather than Evaluation since classification is the main thing carried out in this step.
• Page 7, last paragraph, the reference is missing (Error! Reference source not found is shown).
• Page 9, line 270: The method is being compared not the images.
• Page 9, 3.1.1 Date Description: Date should be changed into Data
• Page 10, line 297: One main problem while using the spectral-based classification approaches is that –usually- they are not able to deal with data with high spatial resolution. This is wisely mentioned here in the manuscript, which is not the proper place. I would suggest addressing/emphasizing this in the Introduction since it is a very important point.
• Page 10, last paragraph: At the end, five or eight features?
• Page 10, Caption of Figure 3: please be consistent with font style/size. Then B:18 should be also bold.
• Page 13, Caption of Figure 4: Why the text is not bold like the caption of the previous figure? Consistency matters.
• Page 13, 3.2.1 Date Description: Again, date should be changed into Data.
• Page 14, the caption of table 8: Bohai Bay should be changed into Baffin Bay data set.
• Page 15, line 474: Again, how to set the parameters in a more hands-off way?
• Page 15, line 476: PCA LayerS
• Page 18, line 533: I think a ‘we’ is missing: “Finally, determine the band ….”

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The concerns by the reviewer have been generally addressed.

Reviewer 3 Report

The issues raised by me have been addressed.  I suggest publishing the article in its current form.