Texture Extraction Techniques for the Classification of Vegetation Species in Hyperspectral Imagery: Bag of Words Approach Based on Superpixels
Round 1
Reviewer 1 Report
In their study, Blanco et al. compare different superpixel-based techniques based on texture metrics to classify high spatial, multispectral UAV data and hyperspectral airborne data in natural and urban landscapes using the support vector machine algorithm. They show that include texture metric classification accuracy in term of overall accuracy is increased. They make a pretty point important point because not too many studies are using texture metrics in this way. Further, they are testing their approach in different settings using different data types. Further, they provide a clear and illustrative workflow diagram. Unfortunately, the manuscript lacks a clear writing style, so methods, results and discussion are mixed. Further, results are not set into context with the other current research, as this section only consists of 5 lines (Ll355-359), which is too short. The manuscript would also benefit from revision by a native speaker, but this is a minor issue.
As a summary, the study is interesting and novel from the technical point of view, but the writing needs profound improvements.
My main comments are as following:
Introduction:
Please point our more the novelty of your Bag of Words Approach.
Finish the section clearly stating out the research question. It sounds too descriptive and detailed.
Methods (“Texture schemes”)
Here, the tense changes abruptly between past and present tense, and sometimes the use of tense does not seem correct. Please check and describe everything you did in past tense.
Using bullet points for subsection is rather unusual. Consider reformatting.
Add a section about the accuracy assessment, e.g. the metrics used, their formulas etc.
Results (“Experimental results”):
L226 ff: Data set description (3.1) is a typical part of the methods. Please move to methods section, and use the past tense where appropriate. This also applies to 3.2.
L294ff “3.3 Results”: This section is not very appealing results section to read, as it includes to much methods material, and because the most important results are not highlighted and set into relation, but the different tables are simply mentioned. Please restructure the results section.
L295: Overall can be misleading depending e.g. on the prevalence of the data, on the target class etc. Please add further accuracy metrics.
Discussion
The discussion directly starts with a detailed description of one single results. Please start with the most important general findings.
Ll329- 354: In these section, the results are described and set into relation without using external references for interpretation. Include these sections in the results part.
L355ff: This is the only section in the discussion that refers to external references. Please extend this discussion with further literature.
My detailed comments are:
L19-20: What do you mean by „ current context of rapid human interventions”? Please reformulate.
L23: “biodiversity identification” is not the proper term here. Please reformulate.
L25: “at lower costs”: Not quite accurate, because you can get satellite-data for free (but a satellite costs more than an UAV, of course). Please reformulate.
L63: “The Random Forest classifier is later used on the resulting 9 bands.” Correct statement, but what does it mean for your study? Please point out the research gap, or use it for discussion or delete.
L60ff: You correctly refer to Feng et al. for a study that used UAV data to map an urban natural landscape. You are also studying natural vegetation. There are recent studies that used UAV data and textural metrics to map natural vegetation such as Oldeland et al. 2017. They also used Random Forest, but also height. Height actually improved the classification accuracy. So please add more references here (Oldeland et al. is a suggestion, please feel free to add further research). You could discuss adding height as a predictor. Was there a reason not to include it?
L97: This is the methods section, right? And the deriving the texture is an important part of the methods, right? Please consider renaming the headings here.
L102: Is superpixel a synonym for segment?
L157: Add a reference PCA as a reduction technique.
L162: Could you add some more details about the SVM, such as how you divided the data set into training and test set, if you optimised any parameter, which software you used etc.? Join with section 3.2.
L165: Can you add a reference for the statement that SVM can handle low number of training samples?
Figure 2: I don’t understand the caption. Can you reformulate it, add some context, maybe formulate it as a sentence?
L310: The fact that you averaged results is an important piece of information, and belongs to the methods section.
L326 ff: “As the standard…”: This belongs to the methods section.
Reference:
Oldeland, J., Große-Stoltenberg, A., Naftal, L., & Strohbach, B. J. (2017). The potential of UAV derived image features for discriminating savannah tree species. In The roles of remote sensing in nature conservation (pp. 183-201). Springer, Cham.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
- How to design the size of the texture window? How did the window being optimized?
- As shown in Figure 1, the image segmentation technology was used in this study. However, both the method of segmentation and the design of parameters for the segmentation were not mentioned in the manuscript.
- The description of SVM classification should be carefully improved, and detailly about the kernel function of SVM classification need to be supplemented.
- The reason for choosing SVM as the classifier is not sufficient. I recommended to compare with other classification methods.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Very good work. I am impressed with authors completion of this valuable task. There are only some comments for the bastract section as all other is ok to me.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The manuscript has been greatly improved. Good job!
