Deep Learning Approaches for the Mapping of Tree Species Diversity in a Tropical Wetland Using Airborne LiDAR and High-Spatial-Resolution Remote Sensing Images

Round 1

Reviewer 1 Report

The introduction gives a full panorama of all remote sensing methods used to map tree species with multi to hyperspectral detectors and LiDAR complementary data including object based technique analyzing local spatial textures.

I just would like to recall that 3-band VHR are only sensitive to pigment variation and texture mainly detectible by shading effects. Hyperspectral usually working from 400 to 2500 nm gives access to other chemical component from water to cellulose and lignin. Therefore, all limitation pointed out by hyperspectral studies using the same deep learning processes are even more applicable to low spectral resolution.

So, the aim of this study is not only to evaluate the potential of deep learning (line 109) it is also to test it on low spectral resolution and high spatial resolution space borne data which is well defined in the title. Forthcoming hyperspectral satellite would certainly enrich this work. Conversely, airborne studies could also implement the proposed deep learning work frame. We always have to compromise between high resolutions (spatial or spectral) but for how long time?

Line 145 What do you mean exactly by dominant species compositions?

Line 159 “we perform data augmentation to the tree samples, the tree patches are rotated, mirrored and flipped randomly” means that the method is sensitive to image orientation. Other technics based on texture analysis can make a difference between texture parameter like anisotropy and orientation (see object based algorithms). So it is not exactly a data augmentation but a procedure required to avoid the distinction of identical textures only differing from each other by orientation changes.

Line 168 RGB images in Tiff format. These are non-calibrated data which may change from date to date with weather conditions, atmosphere transparency and sun elevation changing the texture formed by shadowing length and orientation. Can you discuss that?

Line 172 The DSM is the outer upper envelope of the point cloud. The DTM depend on the classification of tree characterized by scattering of laser returns on different layers of more or less dense leaves in the canopy. Why don’t you used this classification? Then an interpolation of all the points surrounding those clouds of scattered points is used to calculate a model of elevation below the trees. The difference between DSM and DTM therefore gives a volume or height of vegetation. So it is important here to give the parameter used form the tree classification used to calculate the DTM.

Line 178 “our method is designed to infer a patch level prediction” this is unclear. Can you explain?

Line 176 3.1 Overview: is it really useful? For example you introduce here the tree apex without reference but define it latter in section 3.2. Can you reorder?

Line 188 Figure 2. Training samples and data augmentation well exemplify the dependence of your method to sun elevation and orientation. This as to be discussed somewhere in your text. You are note mapping tree species but tree texture pattern. Knowing the corresponding species on the ground of that particular area you can associate them to species. There is no direct identification of species like in other works effectively trying to identify species. There is a nuance between these works which could be discussed. At least it is an indirect tree species mapping. This is also set in the tree classification of the DTM calculation. What was the typical tree crown width searched? This can dramatically change the results.

Line 191 Individual tree detection supposed that each tree present an apex. This is not always the case and a net of trees forming a grove of tree can present only one apex whereas one tree can present a few apexes.  What’s the link with the 64 * 64 pixel of line 205?

Line 229 the 64 * 64 (pixels I guess) image size conduct to an adjustment… can you precise it?

Line 234 all the indexes presented are typically used in biodiversity forest mapping based on field studies. Some references are missing here?

Line 278 The definition of the y value is confusing: “value (the number of trees or the species diversity)”. What’s the link between number of trees and species diversity?  I would imagine that 10 trees can belong to one species or 10 different species. Can you explain?

Line 303 Banana and Papaya have special leaf shapes. Their good score is not surprising.

As the hyperspectral chemical information is missing this work must rely only on texture analysis, which is the main point of this work.

Line 308 Can we have a discussion about de significance of all accuracies presented in Table 1. Some tree type display very bad results. Can we really trust the mean accuracy?

Line 327 4.2 Forest species diversity mapping

The choice of 30m x 30m remains unclear for me. What could be the results with local aggregation of some species in local patches having the same size? The biodiversity would fall down in each 30m x 30m box while it would remain constant over all the area. Can you discuss that?

Line 364 5. Discussion

I am not expert in deep learning architecture so I cannot really evaluate the discussion between algorithms.

On the contrary the point listed in Line 380 to 388 are good ones.

Line 414. Hyperspectral results could also be used to feed up deep learning methods. There is no opposition between methods and many products of hyperspectral analysis could have been used. So I would be more optimistic. However some precisions have to be given in this work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I have reviewed the ms “Deep learning approaches for mapping of tree species diversity in a tropic wetland using airborne LiDAR and high spatial resolution remote sensing images” by Sun et al. The paper reports how remote sensing information potentially can be used to determine biodiversity in terms of forest tree species diversity. The paper is well written and the questions clearly addressed. As a non-expert in the field of remote sensing, I found the paper very technical, and I cannot evaluated the technical quality of the paper. This part of the ms has to be evaluated by someone else. In my opinion, the paper would benefit from being a bit less technical and broaden the general significance of the results. In the conclusion, the authors claim that    conclude that their proposed solution with deep learning approaches is well-suited to forest tree species diversity mapping. In its current for it is actually very difficult to evaluate whether this is true or not. The results of this study is never related to any similar studies so it is very difficult for the reader to understand if this method/approach is any better than other methods. Moreover, the predictive power of the methods is c. 75%. Is this a satisfactory number or not? I would have liked some discussion on what is a reasonable precision, and for what type of assessments such a precision would be useful.  

Related to the precision issue, is it a problem that the most common “tree species” is the lumped class others? The species “others” is in fact more than twice as abundant as any individual species class. I am worried that this actually may have flawed the result, as it makes the fit of the assessed diversity seem a bit better that it actually is. Please clarify how the inclusion of this group may have influenced the results of the study. What would the result had been in terms of accuracy of the methods if the “others” was not included? For a person interested in different measures to describe biodiversity the knowledge that the most common “species” assessed by the method is a lumped group of unidentified species seems a bit unsatisfactory.

 

Specific comments

L 38-46. The forst paragraph can prefereably be omitted

47 support the statement with a reference. Maybe add something on the difficulties/challenges, including high cost associated by assessments of biodiversity. Here, remote senescing methods that can make rough quantification may be very useful.

L.48-49 Here it had been good to provide some examples

L.153-154. Is located the correct word here? Positioned might be better, or?

246. This is an unusual description of Simpson’s div index. Much more common is something like an index that measure diversity, which takes into account the number of species present, as well as the relative abundance of each species. As species richness and evenness increase, so diversity increases.

L.414 What does ”other forest ecological benefits supporting” mean? Please exemplify of delete

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This study applies three deep learning algorithms to map four tree species diversity indices for an urban wetland in Guangzhou, PRC. The algorithm is more of a tree species classification algorithm that then aggregates the classified species to estimate diversity measures. A major concern is that there is no treatment of spatial scale, which is fundamental to understanding diversity patterns. The reference section is noticeably insufficient and makes one think the authors have missed much of the most important literature on the subject. The validation section is vague and focuses on well-known AA equations but doesn't actually explain which ground data was used and how. What was the AA sampling design? In fact, this section is so insufficient that it is difficult to assess the primary conclusions of the paper: that these algorithms are effective in estimating diversity. Finally, the paper was clearly written by a non-native English speaker and needs extensive English editing. That said, the paper is timely and the results are promising. Some specific points follow:

39 - this is a questionable claim

40 - Saarinen et al., 2018 is not a good citation for this. See, for example, Magurran, A. 1988. Ecological diversity and its measurement. Springer, Dordrecht, NL.

40 - "structural variety" is not a common term.

40 - these terms need to be defined

41 - Kimmins, 1997 is not a good citation for this. See, for example, : Staudhammer, C.L. & LeMay, V.M. 2001. Introduction and eval- uation of possible indices of stand structural diversity. Cana- dian Journal of Forest Research 31: 1105–1115.

41 - "one" what?

41-42 - this is not true

47 - this is not true as written. More specificity would help, and a reference.

54-55 - this sentence does not make sense as written

61 - what is "C 4.5"?

66-67 - this sentence lacks sufficient information content

70 - too specific. this is one study, and an old one at that

78-79 - again this lacks sufficient information content. all you said was that it is challenging. this sort off sentence should never be in a journal article

79 - there has been no treatment of spatial scale of diversity nor pixel size. For example see: Hakkenberg, C.R., Zhu, K., Peet, R.K., and C. Song. (2018). Mapping multi-scale vascular plant richness in a forest landscape with integrated LiDAR and hyperspectral remote-sensing. Ecology. 99(2), 474-487.

80 - define texture information

95-97 - this is not a sentence

109-120 - at some point you should mention that you are not using hyperspectral, and state which imagery you are using

110 - "tree species diversity" is ambiguous. Are you mapping "tree species richness"? And at what scales?

118-119 - example of a poorly written passive sentence. The sentence should be in at the active tense. Also this sentence should appear earlier in the paragraph

127 - "semi-natural fruit forest" is not an ecological term

129-132 - this should be in the figure caption

137 - refer to Fig. 1

160-162 - this does not belong in the field sampling section. Also, much more detailed information is needed on the model training process

162-164 - this does not belong in the field sampling section. Also, insufficient information on accuracy assessment process

166 - need far more data on the remotely-sensed imagery. For example where does the RGBs come from? Is it aerial? drone? when was it flown? etc. etc.

167-168 - how was the laser scanner used?

169 - 170 - tile numbers are arbitrary and don't need to be mentioned

173-174 - This is not correct. A CHM does not reduce topographic effects

186-187 - later you mention species richness (238-239). Was this calculated or not?

187 - why did you select these diversity indices?

196-197 - how were the locations derived? what is the accuracy/precision? Does this impact further analyses?

204-205 - insufficient detail on cropping method. how is a patch derived? is it object based around the apex? impossible to review without sufficient information

229 - "64*64" what unit?

233- again, what is an image patch?

235 - do trees "have" species? Doesn't make sense.

239 - 240 - earlier you do not mention species richness

241-261 - reduce and make a table. Each equation needs a reference.

263-264 - unnecessary

272 - yes, but where is the validation data coming from? what is training and what is test? is it internal? external/independent? cross-validated?

276-277 - neither of these equations is necessary because they are very common

292-295 - this should be in Methods

298 - precision or accuracy?

306 - no references should appear in Results

308 - is this all three areas combined?

314 - the first numbers do not look like proportions. Why not sort the results?

316 - "and the dominant tree species are obvious" unnecessary

317-326 - if it is so obvious, then why are you retailing what is obviously in the table in a paragraph format?

331-332 - are these diversity indices correlated?

347 - I don't understand what "the maximum value of field survey within a plot is 9" means

349-353 - this is an opinion and should be in the Discussion.

357 - Why is Merge important for inclusion here?

363 - Why are axes not defined? Also, these should be square plots (x and y axes same range)

365-366 - explain why

391 - based on what evidence?

396 - what about access to training imagery?

397-398 - unintelligible

405 - available?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

This work is now lot easier to read and understand with explanation of the different choices done so I now recommend a publication of it in the present form with a warning about quoting references. I let the final editor work on it.

Author Response

Dear reviewer,

Thank you for taking time to review our manuscript.

We studied your comments and revised our draft accordingly.

Hope that our revised draft will meet with your approval.

 

Best wishes,

Ying Sun

Point-by-point response: 

This work is now lot easier to read and understand with explanation of the different choices done so I now recommend a publication of it in the present form with a warning about quoting references. I let the final editor work on it.

Reply: thank you very much for your positive comments on our article. We modified the references according to the journal, and now the references are cited by bibliography number.  Take the first reference as example, it now reads: “There is much evidence to support the importance of tree species diversity for maintaining wetland ecosystems [1].”

As all the references in our draft were modified, please see details in the main text.

 

Thank you again for all your help.

Author Response File: Author Response.docx

Reviewer 3 Report

I'm happy with the revisions. However, the authors claimed this was checked by a professional, but the ms still clearly contains many grammatical errors. So many, in fact, it is quite distracting to the reader. In fact, beyond grammar, there are still numerous typos (e.g. "filed work" rather than "field work") that makes me think language editing was undertaken carelessly. I think the authors need to make a serious effort to address this concern before considering for publication.

Author Response

Dear reviewer,

Thank you for taking time to review our manuscript.

We studied your comments and revised our draft accordingly.

Hope that our revised draft will meet with your approval.

 

Best wishes,

Ying Sun

Point-by-point response:

I'm happy with the revisions. However, the authors claimed this was checked by a professional, but the ms still clearly contains many grammatical errors. So many, in fact, it is quite distracting to the reader. In fact, beyond grammar, there are still numerous typos (e.g. "filed work" rather than "field work") that makes me think language editing was undertaken carelessly. I think the authors need to make a serious effort to address this concern before considering for publication.

Reply: thank you very much for your positive comments on our article. We corrected the grammatical errors as well as the spelling mistakes as you pointed out. Moreover, we used the editing service of MDPI English Editing for the language editing, and a professional polished the manuscript finally. Please see details in the main text.

 

Thank you again for all your help.

Author Response File: Author Response.docx