Machine Learning Techniques for Tree Species Classification Using Co-Registered LiDAR and Hyperspectral Data
Round 1
Reviewer 1 Report
The manuscript considers main tree species classification using remotely sensed data as an input in machine learning based classifiers. The manuscript is well written and the topic is important. The presented idea nor the results are particularly novel. However, the manuscript has its own merits, which could be enhanced to make the paper recommended for publication. This requires improvements in Materials and methods section and extensions in analysis.
General comments
Introduction
Reader gets an idea that the methods are developed and tested are ITD (individual tree detection) methods. Later, it comes clear that the analysis is done at the plot level. This should be made clear in introduction and in abstract. Also, it should be made clear why analysis was done at plot level and not at tree level? This comes clear after one has read the whole manuscript, but it is not explicitly told.
What is the essential difference of co-registration and data fusion between the data set used here and, for example, in the following two research papers?
1. Packalén, P., Suvanto, A., and Maltamo, M. 2009. A two stage method to estimate speciesspecific growing stock by combining ALS data and aerial photographs of known orientation parameters. Photogrammetric Engineering and Remote Sensing 75(12): 1451-1460.
2. Valbuena, Ruben & Mauro, Francisco & José Arjonilla, Francisco & Manzanera, J. (2011). Comparing airborne laser scanning-imagery fusion methods based on geometric accuracy in forested areas. Remote Sensing of Environment. 115. 1942-1954. 10.1016/j.rse.2011.03.017.
Materials and methods
The remotely sensed data and field data are not described with enough details. It is unclear how the lidar variables were extracted. Also, field data description is incomplete. Figure of a map showing, for example, one field plot with remotely sensed materials and field measured trees would be very beneficial for a reader to get an idea of the materials. Also, see specific comments for details.
The variable selection methods, or dimensionality reduction, tested were only two, PCA and the method based on classification tree breaks. It would improve the scientific merit of the manuscript if more methods were tested. See, for example:
1. Packalén, P., Temesgen, H. and Maltamo, M. 2012. Variable Selection Strategies for Nearest Neighbor Imputation Methods Used in Remote Sensing-Based Forest Inventory. Canadian Journal of Remote Sensing 38(5): 1-13.
Specific comments
Line 110, Abstract: define acronym LiDAR
Line 30, Keywords: Many of the keywords are mentioned in title. Usually keywords should be something complementary to title.
Line 33: Define abbreviation “GIS”.
Lines 43-45: The known problem in individual tree detection (ITD) methods is that height and diameter can be estimated for detected trees only. It is not simple at all to valuate the timber based on ITD because the tree detection is not correct. Or at least it is not simple if the aim is to get precise and accurate results. Authors should mention that the species identification is not the only challenge of lidar based ITD methods.
Lines 47-49: I do not understand. How sufficient point density improves the signal of individual lidar data points?
Lines 55-56: Would full waveform lidar be superior, in this sense? Discrete return lidar systems may have echo form canopy top, ground and few observations between.
Lines 149-150: It is not clear what is the spatial unit for which the lidar features were extracted? For 13 m2 pixels, or for 25 x 25 m subplots? And how? From normalized point clouds or from some pixel data processed earlier?
Line 181: Orange? Please define.
Lines 194-196. I do not understand how cross-validation resampling is applied in this case. Please describe with more details in text.
Lines 230-236: You have examined DBH distribution only. It would be more informative to compare other characteristics, too (height, volume). And more importantly, since the analysis and classification is done at the plot level, not at individual tree level, you should compare plot level statistics, not individual tree level statistics.
Line 361: Define SVM when support vector machines is mentioned the first time in text.
Lines 214-216. Can you claim that the best machine learning algorithms based on full list of explanatory variables are the best ones when using reduced lists, too? I think you should do the data reduction analysis for all machine learning methods.
Author Response
Thank you for your comments. Please see the attached file for our line-by-line responses.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Evaluation
Overall evaluation
The paper presents an analysis of the information contained in simultaneously acquired discrete return airborne laser scanning data and hyperspectral images with regards to species identification. The data is used jointly to identify the dominant species in 25 m x 25 m plots (and larger ones) within two small experimental forests of the northeastern USA. The data was acquired with the G-LiHT system from NASA. The focus of the paper is on feature selection (dimensionality reduction), classification methods (6 different ones are tested), and comparison of the respective and combined utility of the data components (lidar, reflectance, vegetation indices derived from the imagery, and combination of these three). Unsurprisingly, the best species identification accuracy is achieved with the combined datasets (68% in one scenario). What is notable is the large number of species (18 dominant species in total over the 2 sites, with a maximum of 13 for one of the sites), which is well above the average of other papers on the subject. As could be expected also, the accuracy is lower for the site with more species (13), compared to the other (9).
This is not the first time that airborne lidar and hyperspectral data are combined. The G-LiHT system has been around for a while, at least since the datasets reported in this paper were acquired (2012). So in my view, the originality stands more in the dimensionality reduction and use of different classification approaches, although this has also been done before. See for example Ghosh et al (2014), which is surprisingly the 69th reference of the paper, appearing in the discussion, but not in the introduction, despite being a similar study.
The manuscript takes quite a few shortcuts, in the literature review and in the methods. Some choices of 3D features are questionable (see detailed remarks section). Due to its modestly innovative character, and lack of methodological details, the scientific contribution made though this paper is relatively minor. Before being accepted, several modifications should be made, and the paper re-reviewed.
Detailed remarks
1. Introduction
In general, the introduction is not too well focused on the main subject: combination of low density lidar to somewhat low resolution hyperspectral airborne imagery (1 m is rather low for airborne platforms). The discussion about branching pattern (which require a point density of 20 or more points per m2 in my view) leads the reader to think the paper is going to be about very high density point clouds. But actually, the lidar resolution is 3.6 m x 3.6 m (assuming the 13 m2 pixels are square). The typical lidar resolution is more around 50 cm. So the narrative in the introduction should reflect the focus of this study, and leave individual trees, branch overlap and bifurcation ratios alone.
LL48-9 / "few studies have... sufficient point density" What would be the required point density? What is low or high density? What is the barrier?
L51 / What is ecosystem "identity"?
L55 / "Thus, discrete-return LiDAR data may offer valuable insight into branch overlap ". High-resolution full waveform lidar (as opposed to discrete return) would do a much better job at this.
L71 / There are at least 3 dozens peer-reviewed studies on using lidar for tree species identification since 2009. You here cite just 3, dating from 2007-2009. Two of these were not published in journals but as conference papers. You have not shown that you know what has been accomplished so far. This needs a substantial upgrade.
LL82-7 / I think it is worth mentioning multispectral lidar at this point. I can think of at least three studies using MS lidar for species identification (not to mention hyperspectral lidar).
LL114-7 / This is a very specific problem. It reads as a part of the methods section. The general problem is feature selection to avoid the curse of dimensionality, whether with G-LiHT, or in other much bigger forests.
L119 / What is "filtering" in the context of reduction of feature dimensions?
LL125-6 / What is the gap? There is plenty of literature on all the aspects discussed in the introduction. So, what is it specifically that is missing?
2. Material and methods
L134/ The points were aggregated to 13 m2 pixels. But what was the native point density? Have you aggregated two points per pixel, hundreds of points per pixel? Why such a resolution disparity between the lidar and hyperspectral pixels? Why not aggregate the lidar data to 1 m2 pixels? Puzzling.
LL139-140 / Acres? Did hectares fall out of fashion? Also, don't put a hyphen between the numerical value and the unit (e.g., 1-m2 on line 134, 558-acres here).
LL142-3 / How many plots were there on each site?
L144 / Don't mention variables that are not used. Biomass is not used in this study.
Table 1 / There is a missing vertical bar in the equation of the AAD. In the equation of CCR, use "/" instead of ":", as you did for D0-D9 and VDR. "Fraction of all returns intercepted by trees": returns are not intercepted by trees, laser pulses are. The "Mean" parameter will be influenced by the size of the tree. This means that species identification will be influenced by the height distribution within the training sample (of which we don’t know the size). The result of this is that you will not know if you are able to distinguish a species because of the architecture of the trees and its spectral signature, or simply because the trees of that species in your sample had a special size (e.g., because they were younger, etc.). The fact that you are using this makes the results impossible to generalize to other regions. Either remove this feature from the analysis or discuss this issue in the discussion section.
Table 2 / Replace "Citation" by "Reference".
L163 / Replace "shapefile" by "polygon".
LL163-4 / "The species with the greatest number of individual trees (stem count) was chosen as the dominant species ". A very questionable choice. The logical way would be to choose the species with the largest plot-wise basal area. What is the impact of this on the results? What if there are numerous small trees that don't reach the top of the canopy? That would be your dominant species even if there is no chance you see them from the air? Using basal area as a criteria might actually improve your results.
L166 / What is the disturbance history of those forests? They will affect the age of the different stands, and therefore, the relationship between height and species (in addition to the characteristic height at maturity that varies with species).
L178 / What does "resampling" refer to in the context of classification?
3. Results
LL232-3 / "DBH values are not sufficiently informative to classify tree species by this metric alone": why would we expect this?
LL234-5 / "Nonetheless, the weak trends that can be detected suggest that further examination of tree structural information using LiDAR metrics is sound avenue of analysis": this is why hundreds of papers on that subject have been published in the last 20 years. In this context, the statement sounds quite naive.
L295 / One word is misspelled.
What could be the impact of canopy gaps on the results?
4. Discussion
LL326-328 / One word is missing in this sentence. Additionally, 500 pixels is either 500 m2 (hyperspectral), or 6500 m2 (lidar). Your plots cover 10000 m2, and the subplots 625 m2. Where does the 500 pixel figure come from? I see no correspondence.
L335 / What about the contaminating effect of the shrub understory, or bare ground?
L338-9 / I did not see such a statement in reference 18. Is this an error of referencing? If not, can you point in your reply where in that paper such a statement is being made? I could not find it.
L379 / Gaps are revealed by lidar alone. Data fusion is not necessary.
L381 / A word is missing.
Reference
Ghosh, A.; Fassnacht, F.E.; Joshi, P.K.; Koch, B. A framework for mapping tree species combining hyperspectral and LiDAR data: Role of selected classifiers and sensor across three spatial scales. Int. J. Appl. Earth Obs. Geoinf. 2014, 26, 49–63.
Author Response
Thank you for your comments. Please see the attached file for our line-by-line responses.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The manuscript has improved significantly from the previous version. The description of the methodology is clearer. There are still some issues that should be improved.
Lines 84-89 are in contradiction to what is said in lines 46-48. In lines 46-48 it is said “… accurately… … each tree…” and in Lines 89-95 it is told that ITD does not actually work in practice. Please correct the sentence in lines 46-48 so that it is not in contradiction to what is said later.
Authors should add chapter about variable selection methods in discussion. Only two methods were tested, PCA and the method based on classification tree breaks. I understand it is quote laborious to add more methods at this point, and it may not be even necessary. However, the issue of variable selection should be discussed. See for example:
1. Packalén, P., Temesgen, H. and Maltamo, M. 2012. Variable Selection Strategies for Nearest Neighbor Imputation Methods Used in Remote Sensing-Based Forest Inventory. Canadian Journal of Remote Sensing 38(5): 1-13.
Lines 406-408: Is there a mistake? K-nn, Neural networks and RF were the best according to the results and here you mention SVM and CN2 rulers and not RF and k-nn.
Lines 410-412. New information. Should be mentioned in discussion or even earlier.
Author Response
We appreciate your comments on this manuscript. They have provided substantial contributions to the content of this work. Please find attached our responses.
Author Response File: Author Response.docx
Reviewer 2 Report
The authors have made a very good job at modifying the paper according to my comments. Most changes were done, but in their reply, they have skipped my general comments and the first comment on the introduction:
“In general, the introduction is not too well focused on the main subject: combination of low density lidar to somewhat low resolution hyperspectral airborne imagery (1 m is rather low for airborne platforms). The discussion about branching pattern (which require a point density of 20 or more points per m2 in my view) leads the reader to think the paper is going to be about very high density point clouds. But actually, the lidar resolution is 3.6 m x 3.6 m (assuming the 13 m2 pixels are square). The typical lidar resolution is more around 50 cm. So the narrative in the introduction should reflect the focus of this study, and leave individual trees, branch overlap and bifurcation ratios alone.”
In the introduction, there is still a long section about branching patterns (LL54-71). Tree species differ in the overall shape and proportions of trees, foliage type (broadleaves vs. needles), foliage reflectance, LAI, etc. I don’t understand why the authors choose to focus on branching pattern, especially as the lidar data they are using does not have a density necessary to resolve individual branches. This should be adjusted as it still leads the reader to think this is going to be a study about characterizing branching patterns from a reconstruction of individual branches.
Author Response
We appreciate your comments on this manuscript. They have provided substantial contributions to the content of this work. Please find attached our responses.
Author Response File: Author Response.docx
