Fine Classification of Urban Tree Species Based on UAV-Based RGB Imagery and LiDAR Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Improving the success of tree species identification remotely is an
important contribution to science.  The use of UAV is a very economical
approach to the success of this venture and thus this article is an
amazing contribution to improvement in scientific literature.

But this is one test of one set of tree types.  If this is applied to
the same trees in a different area, or to different trees, would the
same results occur?  Thus the lack of redundancy to the experiment
significantly questions its scientific validity.  Testing it at
a different time of year was outstanding.

I feel the English corrections would only be resolved by a person
with expertise in the subject (me), not an expert in English.

(1) Line 47-48 - lack of 3-D is not specific to urban environment
(2) Line 61 define "fine" tree species classification
(3) line 62 take out "the provision of" - define "high" as opposed
to other resolutions
(4) line 78 - why are there limitations?
(5) line 82 - "exact" incorrect - Lidar has error
(6) line 85 - But photogrammetry can determine 3D?
(7) line 89 - DAR should be dar
(8) line 95-96 limited studies?  Reference them.
(9) line 100 merge not merger
(10) line 105 references for salt and pepper
(11) line 110 reference Plurality Filling
(12) line 126 - 696.5 m annual precip is almost 2 m/ day - impossible
(13) Figure 1 - text in figure is not readable
(14) Figure 2 - the lidar images are not readable
(15) Line 146 - "at" should be "on"
(16) Line 155 - 5 cm accuracy - how was this determined
(17) Line 157 - define noise
(18) Line 168 - "precise" and "hand held" have different meanings
mostly define precise
(19) Line 195-196 - grassland cannot be 2 m high
(20) Line 210 "by Smith" needs a reference in text in addition to table
(21) Line 218 - no reference
(22) Line 277 - continued use of "precise" and "accurate: but it is not defined
(23) Figure 6 - unreadable text
(24) Line 337 - producer is produced
(25) Figure 7 - unreadable text
(26) Figure 9 - unreadable text
(27) Line 431 etc. - this cannot be generically applied as tree type, size, and flying
height/ground resolution will change results (probably)
(28) Line 477 - "excessive" becomes "additional"
(29) Line 478-479 - sentence structure needs work
(30) Line 493 "more excellent performance"needs English work
(31) Line 505 take out "been"
(32) Line 518 "excessive involvement" - what does that mean?
(33) Line 521 - take out "The"
(34) Figure 11 - labels on plots not readable
(35) Line 544 - change "mainly to explore" to "explored"

Comments on the Quality of English Language

See above

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper tests the capacity of UAV acquired RGB and LiDAR data for 10 species identification in urban environment.

The methodology is not new, except perhaps for HSV features. The paper proposes 7 types of RGB and LiDAR features. In this paper 7 different experiments were tested, with different feature type selection. The random forest algorithm was used for classification. 2 post processing algorithms were tested (window sliding and plurality filling). 7 different resolutions were tested, between 0,1 m to 2 m.

The English is comprehensive and overall, well written.

Introduction is well presented.

The use of lidar in the paper is deceiving. For the quality of the lidar data acquired by the drone, the authors could extract more valuable information than only the nDSM and intensity features. First, the nDSM feature representing the tree height, cannot be reasonably used for tree species classification. The trees of any species can have different ages and heights. The nDSM feature can be used in a classification model to stratify the sample according to the tree height. This way, the changes in tree morphology according to the tree height, and by consequence influence the RGB and LiDAR features, could be addressed.

As you used the nDSM, equivalent of canopy height, it is mandatory to provide a table with statistics of tree height by species (max, min, mean, coefficient of variation).

Do you calculate Lidar features from LiDAR point cloud? There are other features largely used in the literature and better suited for tree species classification, like the height percentiles (normalized by the tree height), vertical distribution of points, geometry features that characterize tree shape.

Line 39: “its function” refers to what? Is it the same as the “importance[s] mentioned line 38?

Line 42: What is included in “traditional method”. Please, be more specific to what type of species classification you refer to.

Line 49: Give more details on what “traditional satellite remote sensing” is : range of image resolution, type of satellite, wavelengths.

Line 53: “us” give the impression that you are describing here your methodology. Instead, you in the Introduction section where you should talk about the literature review.  Erase “us”.

Line 55: What is the meaning of “various dimensions”?

Line 100:  merge

Line 110: The “plurality filling” is a method developed in this paper. Is this the novelty of this paper. Can you give a reference for this algorithm?

Line 126: “ precipitation of 696.5 m” seems like the Great Flood with near a mean of 2 meters by day… The precipitation unit should be “mm”.

Figure 1 : The text for coordinates, scalebar and legend are too small and cannot be read.

Line 150: “Hasselblad camera” precise the model

Table 1: The number of pixels by species is not useful information, rather give the percentage of total area. But we need more information concerning mean tree height by species, min and max of tree height, crown diameter.

Figure 3 and other occurrences of feature combination:

-        Do not understand why HSV is mentioned twice in the same model

-        Do not understand why GLCM is mentioned twice, with negative and positive signs

-        There is no perfect correspondence between features in the second (large) box and the third box of Feature combination. Specify the abbreviations in parenthesis after each of the feature classes in the second box (ex: “Vegetation indexes (DEVI)” – and… is this the correct correspondance?)

-        You compared 2 post processings. Why they are not mentioned both in the Figure 3?

Figure 1 : The multisegmentation is specific for each model? Do you calculate separately 6 different segmentations from each of the 6 feature combinations? Or only from RGB+DEVI+ nDSM+ GLCM?

Line 183: Find another abbreviation for “HSV-GLCM” because it is confusing when writing the feature combination. Suggestion : “HSV_GLCM”, “HSVGLCM”, or at least put spaces before and after the “+” signs.

Section 2.3.1. The result of “space extraction of urban trees is all surface covered by trees? Did you do a segmentation by individual tree crown? Do you use the manual delineation only to compare. Why you don’t use directly the manually delineated trees? The multi-scale segmentation resulted in an over-segmentation. How did you relate the manually delineated tree crown to the corresponding segments obtained automatically?

Line 210: “Smith” needs a reference

Lines 229-236: Why just 2 lidar variables, when you have such a dense lidar point cloud and ha multiple feature possibilities that are more adequate for species identification? The nDSM is related to tree height. How the NDSM could help in species identification without introducing a bias related to the tree height/age?

Line 287: Why you have 7 experiments here and only 6 in the Figure 3? In Figure 3 is the little left box an experiment too? Why is he different from the others? Please, put numbers on the experiments in this figure and represent them with similar boxes.

Line 288: What is a multi-dimensional feature?

Line 292: Why did you go directly from the first experiment to the last one?

Lines 294-296 : introduce these lines just after line 291.

The column in Table 3 is repeated in Tables 4 and 5. So Table 3 can be omitted.

Table 3: What includes the “RGB” only the 3 bands of the image or also the indexes calculated in the Table 2? Why the index “DEVI” is counted separately in the  7 feature combinations? How do you choose this index? Why not the other 6 indexes? If you do not use them, why did you calculate them. A Table with all features arranged by their type is needed for understanding how many features there are mentioned by each element (ex. RBC, GLCM, HSV).

Figure 6: Interesting representation. However, outside numbers are too small.

Lines 355-356: Explain the method. What is the “segmentation scale”, the “shape factor” and the “pactness factor”. What do represent the corresponding numbers. What software did you use to do this segmentation?

Line 357: The over-segmentation is compared to the manually delineated trees ? Can you develop on that? Some statistics. How many segments per tree? All species are over segmented equally? Are difference in segmentation depending on species?

Figure 7 : Legend too small.

Line 376: The “Experiment 5, utilizing  only a single vegetation index”- from the feature combination of this experiment I have a difficulty to state why there is only one index. Until this sentence I didn’t remark the “VIs” abbreviation. Please make clear in the text, in the Figure 3 and in a separate Table the list of abbreviations, the corresponding description of the feature classes and the specific description and abbreviation of all features in each feature class.

Lines 383-384 and 494-497 are methodology and need to appear in the Methodology section, integrated or near the section 2.3.4.

Figure 9 (b) “post-window sliding result of the optimal scheme” is not clear at what it refers until now. Need to specify that in the methodology.

Figure 9 : numbers are too small.

Lines 432-433: the resample method needs to be specified in the methodology section. At this point it is not clear how did you do this. Majority, max, median, mean value, bilinear interpolation, other algorithm?

Line 450: Instead of the “Table 7” there is : “Table 1. This is a table. Tables should be placed in the main text near to the first time they are cited.” Replace the Table number and title with the appropriate information instead of the instructions from the template. Moreover, for the content of the table, you repeated exactly the Table 5. Please replace with the appropriate results for Table 7.

Line 458: Erase “of”

Line 524: erase “that of”

Figure: 11 and 12: Text for numbers and species on the plot’ axes is too small.

Line 499: How the over segmentation fits or not the tree crown? What are the consequences of this?

Line 515: It is not clear what are the “additional vegetation indices” that “were added”.

Lines 506: Why the second dataset acquired in October is not mentioned in the methodology? Why the results of this comparison appears only in the discussion and not in the results section. Is this a part of your study or it compares with data from another study?

Line 571: What is the meaning of this statement : “No new data were created or analyzed in this study.”? You didn’t collect data (Lidar and RGB)? You didn’t analyze them? Or you don’t share them?

Comments on the Quality of English Language

English is correct. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

It's an interesting job, the theme is relevant, especially for researchers and decision-makers on the issue of urban vegetation.

The methodology is well detailed and gives us a good idea of the executed project.

The results are well discussed, the data are supported with tables.

The research has a great possibility to be replicated worldwide.

Two comments that the authors may judge whether necessary to incorporate into the article.

• The study relies exclusively on one classifier, Random Forest. Why weren't other classifiers tested? In the future, it would be interesting to explore the use of deep learning methods.
• I believe that the passage that addresses the combination of Lidar data and raster images is very synthetic. For me it was difficult. I believe that it is an important point in the study and may be a bit complex and not totally clear in the text. Considering the theme that can be replicated worldwide, I suggest you expand this stage with more details.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors put the effort to carefully address each of my comments. The overall manuscript is improved, especially the field data description and the methodology. The quality of the figures is improved too.

 

Response 2 : Indeed, we had previously attempted to extract height percentile data and intensity percentile data from LiDAR point clouds. However, due to issues with data quality, we encountered a significant number of invalid values, rendering the height and intensity data unusable.

Upon closer examination, we identified a prevailing issue of tree occlusion within the study area, potentially affecting the quality of LiDAR point cloud data obtained from overhead sources. Consequently, we made the decision to focus solely on two LiDAR features throughout our study: one for distinguishing different tree species based on height and the other providing intensity information for classification purposes.

Response 19 : In fact, we previously attempted to extract height percentile data and intensity percentile data from the LiDAR point cloud. However, due to issues with data quality, we encountered a significant number of invalid values, rendering the height and intensity data unusable. Upon analysis, we identified a prominent issue of tree occlusion within the study area, which may have affected the quality of LiDAR point cloud data acquired from aerial sources. As a result, we ultimately only utilized two LiDAR features throughout the study: one for distinguishing between different tree species based on height and another for providing intensity information for classification.

The information in responses 2 and 19 (previous lines) needs to appear in the paper, not only in comments for the revisor. The problem of the impossibility to calculate largely used lidar features for small tree crowns, or, in your case, for too small crown segments, is important to mention. It is also the main reason for the lidar feature choice in this study, only the nDSM and intensity features. You need to explain the combination of point density and average segment dimensions on the possibility/impossibility to calculate certain features, especially those based on statistics, that need more than 3 points per segment. Mention also the impossibility of random forest to compute a classification with databases presenting No Data for classification features.

Response 19: Nice boxplot, it helps visualize the tree height distribution and can facilitate the comprehension and the visualization of the structure of the tree sample. I am aware it repeats information now available in Table 3, but it can be helpful to have in the paper. My observation concerning the tree height (or the use of nDSM) is that the tree height is not balanced between species in the tree sample. So, there are little trees, especially MS and JC, and, on the other hand, there are taller trees, such as PO and PP. To distinguish between these two categories, the tree height (nDSM) would perform very well. However, when applying this model to another region where there will be smaller PO and PP and higher MS and JC, the model will not perform well. If there are species that grow higher than the others, the height could be a good feature to distinguish these species, but only for mature trees while it will not recognize the smaller trees of the same species. I am interested in what was the feature importance ranking for the best model. What was the nDSM feature ranking in this model?

 

Point 3: Line 39:

“Here, "function" refers to the role of urban forests”

If “its” replaces the words “forests” it should be “their” instead, as it is a plural noun : “their function”.

Lines 267-278: The title of the section 2.3.4. is “Post-classification processing”, however, the first paragraph concerns the pre-classification processing. This paragraph should be put before the classification paragraph (2.3.3).

“Data Availability Statement: No new data were created or analyzed in this study. Data sharing is not applicable to this article” This statement contradicts what you mentioned in the response to the revisors, because you created new data and it is not only a theoretical article, but data cannot be shared. I think the more appropriate statement should be : “Restrictions apply to the datasets.”

References : When not all authors are mentioned you put “etc.”, abbreviation from the latin “et cetera” which means “and other similar things”. Usually in references it is used “et al.”, abbreviation from the latin “et alia” which means “and the others”.

Line 697: Reference “Meyer, G.E.; Hindman, T.W.; Laksmi, K. In Machine vision detection parameters for plant species identification, Bellingham WA, 1999/1/1 697 1999; pp. 327-335.” : If this is the same conference paper “George E. Meyer, Timothy W. Hindman, and Koppolu Laksmi "Machine vision detection parameters for plant species identification", Proc. SPIE 3543, Precision Agriculture and Biological Quality, (14 January 1999); https://doi.org/10.1117/12.336896 » the conference title for the proceedings need to be specified.

 

 

 

 

Comments on the Quality of English Language

The English is correct. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf