Round 1

Reviewer 1 Report

McPartland et al. improved their manuscript in terms of structure and clarity.

In addition, the authors took into account all the comments of the different proofreaders, and have made a special effort to relaunch some of the experiences (Random Forest in particular), better justify their assertion (test for small data sizes) or to correct experimental design errors (in the end, it is not a question of constrained PCA but of Constrained Correspondence Analysis). All these efforts and the rigour of the authors must be highlighted.

To go even further, it would be a great help for research if the authors could provide their code on a Gitlab-type platform to allow reproducible science.

Author Response

McPartland et al. improved their manuscript in terms of structure and clarity.

In addition, the authors took into account all the comments of the different proofreaders, and have made a special effort to relaunch some of the experiences (Random Forest in particular), better justify their assertion (test for small data sizes) or to correct experimental design errors (in the end, it is not a question of constrained PCA but of Constrained Correspondence Analysis). All these efforts and the rigour of the authors must be highlighted.

M. McPartland – Thank you!

To go even further, it would be a great help for research if the authors could provide their code on a Gitlab-type platform to allow reproducible science.

M. McPartland – I will take the recommendation to put the code for this project in a publicly-accessible platform under advisement. I am certainly committed to reproducibility, and would like to share the code the I’ve developed more broadly. Although I do not have the time to create a Github for my code before this revisions is due back to the editors, I will make the commitment to make my code available in the coming months. 

Reviewer 2 Report

 @page { size: 8.5in 11in; margin: 0.79in } p { margin-bottom: 0.1in; line-height: 115%; background: transparent } a:link { color: #000080; so-language: zxx; text-decoration: underline }

The manuscript of McPartland et al. have improved greatlely seen the first revision. At this point I only have minor comments:

L111-113: In which sense? How do vascular plant species diversity relate to peatland functioning? There is a nice new paper (see below) which summarizes the interconnections between peatland diversity and other plant properties with carbon storage. This could add also some background in the Discussion about how the relations between vegetation properties would change under a climate change scenario.

Lopatin, J., Kattenborn, T., Galleguillos, M., Perez-quezada, J.F., Schmidtlein, S., 2019. Using aboveground vegetation attributes as proxies for mapping peatland belowground carbon stocks. Remote Sens. Environ. 231, 111217. https://doi.org/10.1016/j.rse.2019.111217

L297: ‘we performed...’

L498-499: Add the reference provided in previous comment here. Furthermore, a few lines of how these functions would change would be nice. For example, a shift in diversity due to climate change will affect productivity and carbon storage how?

L499: Tilman et al. 1997).2012

L504-505: None of these references did studies in peatlands. Better provide examples of studies predicting vascular plant species diversity specifically your vegetation type. E.g.:

Cabezas, J., Galleguillos, M., Perez-Quezada, J.F., 2016. Predicting Vascular Plant Richness in a Heterogeneous Wetland Using Spectral and Textural Features and a Random Forest Algorithm. IEEE Geosci. Remote Sens. Lett. 13, 646–650. https://doi.org/10.1109/LGRS.2016.2532743

Monique Poulin , Denis Careau , Line Rochefort , and André Desrochers, 2002. From Satellite Imagery to Peatland Vegetation Diversity: How Reliable Are Habitat Maps?. Conservation Ecology 6, 16.

Author Response

The manuscript of McPartland et al. have improved greatly seen the first revision. At this point I only have minor comments:

L111-113: In which sense? How do vascular plant species diversity relate to peatland functioning? There is a nice new paper (see below) which summarizes the interconnections between peatland diversity and other plant properties with carbon storage. This could add also some background in the Discussion about how the relations between vegetation properties would change under a climate change scenario.

M. McPartland – I added in a couple of citations and an additional sentence linking diversity with ecosystem processes (Line 111-113)

Lopatin, J., Kattenborn, T., Galleguillos, M., Perez-quezada, J.F., Schmidtlein, S., 2019. Using aboveground vegetation attributes as proxies for mapping peatland belowground carbon stocks. Remote Sens. Environ. 231, 111217. https://doi.org/10.1016/j.rse.2019.111217

M. McPartland -  Citation added (Line 504)

L297: ‘we performed...’

M. McPartland – edit made

L498-499: Add the reference provided in previous comment here. Furthermore, a few lines of how these functions would change would be nice. For example, a shift in diversity due to climate change will affect productivity and carbon storage how?

M. McPartland – Citation added (Line 504). Additional phrase of clarification added about the relationship between diversity and ecosystem function (line 498).  

L499: Tilman et al. 1997).2012

M. McPartland – This is the citation that we are referencing here:

Tilman, D., Knops, J., Wedin, D., Reich, P., Ritchie, M., & Siemann, E. (1997). The Influence of Functional Diversity and Composition on Ecosystem Processes. Science, 277(5330), 1300–1302. https://doi.org/10.1126/science.277.5330.1300

L504-505: None of these references did studies in peatlands. Better provide examples of studies predicting vascular plant species diversity specifically your vegetation type. E.g.:

M. McPartland – citations added, and language edited (line 504).

Cabezas, J., Galleguillos, M., Perez-Quezada, J.F., 2016. Predicting Vascular Plant Richness in a Heterogeneous Wetland Using Spectral and Textural Features and a Random Forest Algorithm. IEEE Geosci. Remote Sens. Lett. 13, 646–650. https://doi.org/10.1109/LGRS.2016.2532743

Monique Poulin , Denis Careau , Line Rochefort , and André Desrochers, 2002. From Satellite Imagery to Peatland Vegetation Diversity: How Reliable Are Habitat Maps?. Conservation Ecology 6, 16.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

1. General impression

The paper puts into perspective topics that will be increasingly highlighted because of environmental problems. What can be done to link ecological problems with remote sensing techniques that allow for long-term monitoring? (EBV's should be mentionned) The difficulty of the subject and the challenge raised must be highlighted: peat bogs have highly heterogeneous plant communities and the spatial scales used in hyperspectral remote sensing are not necessarily adapted.

The manuscript should be published after some modifications. Link between spectral signatures and ecological properties/diversity would be appreciated. Moreover, there is a lack of data on the percentage of covera or the different Shannon diversities for example. However this study show encouraging results.

2. Abstract

l34-37: can the authors quantify ``strong effect'' and ``positive relationship'' ?

3. Introduction

A detailed plan at the end of the introduction would be appreciated.

4. Methods

4.1. Study Sites

A first map to locate the sites would help. The presentation of the data is a little confusing:

APEX consists of 3 water manipulation: it is not clear that these treatments are studied or used later in the article.

SPRUCE: in total, ``we tracked 33 vegetation community plots…'' if this is the case, it should be mentioned in the following section.

4.2. Data Collection

4.2.1. Vegetation Cover Sampling

In addition to Table 1, a summary table of the different used data would provide a better understanding: e.g. APEX, number of plots per treatment, percent cover,…

l188: how many plots ? map to locate transects ?

It is regrettable not to have a table of floristic records for SPRUCE. Could it be added as supplementary materials ?

As for APEX, a map to locate the plots would be appreciated.

Should it be possible to have a summary table with the dominant species or the PFT depending on the plots for the 2 study sites?

4.2.2. Spectral Sampling

Could the authors give instrumental characteristics of the spectroradiometer (spectral resolution and spectral sampling according to the spectral range)?

figure1: Wavelength instead of Bands

l218: \SI{44}{\centi\metre}

Eq. (1): \[2 \times \textrm{height} \times \textrm{tan}\left( \dfrac{25\si{\degree}}{2} \right)\]

How many samples were acquired during one measurement ?

l226: Did the three measurments performed on the same plot or ``around'' to capture vegetation variability?

l224: Did the measurements take place over several days? If so, it should be mentioned.

l227-228: did the authors use an atmospheric transmittance model to remove the bands ? if so, which one ?

l233: As for the previous instrument (ASD), could the authors provide instrumental specifications ?

As the acquisition dates differ and the sites are not at the same latitude and longitude, it should be ensured that the position of the sun does not have an impact on the measurements. The authors should specify the position of the Sun (azimuth angle).

l255: Did the measurements take place over several days? If so, it should be mentioned.

4.2.3. Aerial hyperspectral data collection

Did the hyperspectral data collected only on APEX site ? Did the imagery collected on June of 2015 or July of 2014 ?

l266: Could the authors provide the instrumental characteristics of the hyperspectral camera ? Is the camera composed of two captors ? In that case, spectral resolution and spectral sampling are not the same in the VNIR and the SWIR. Besides, are the authors sure of the altitude of the flight ? \SI{200}{\metre}a.g.l. ?

l272: could the authors precise the plugin ? Semi automatic classification ? dzetsaka ?

4.3. Data analysis

4.3.1. Spectral data analysis

l281: analysis

Could the author add reference about constrained principal component analysis ?

l290: is alpha-value different of p-value (which is usually used)?

l302: ANOVA

l312: are there enough points to confirm normal distribution ? I don't understand what is normally distributed: spectral coefficient of variations, spectral heterogeneity, community heterogeneity per PFT ? per species ? all data ? Could the authors provide the diagnostic plots ?

4.3.2. Hyperspectral image analysis

l321: could the authors convert into meters ?

l323: which software was used to mosaic and georeference images ? which parameters were used for RF (number of estimators or trees, depth of each tree, …)?

5. Results

5.1. Community drivers of spectral response

Could the authors show the different spectral signatures ?

The relationship between PCA and explained variables is not clear. Can the author detail how variables can be explained by reducing spectral dimension ?

5.2. Relationships between species diversity and spectral variation

f_11 means F-score ? if so, authors should mention it or detail it.

Can the "not difference" seen be explained by different location of sites and different dates of measurments and different "conditionment" ?

5.3. Hyperspectral image analysis – mapping of PTFs

l405 - figure 8: how about ground truth to compare ?

RF is sensitive to the ratio of sampling the individuals in population and parameters. This is why it is extremely interesting to specify the number of individuals as well as the parameters inherent to the algorithm.

6. Discussion

6.1. Hyperspectral remote sensing of peatland to experimentally-mainpulated climage change

It is difficult to conclude because the sites do not benefit from the same acquisition protocols. Moreover, the number of individuals in the different samples does not seem sufficient to conclude (given the large number of hyperspectral parameters in terms of wavelengths).

6.2. Remote sensing of boreal peatland species diversity

Authors should either mentionned:

  Erudel, T., S. Fabre, T. Houet, F. Mazier, and X. Briottet (2017). “Criteria Comparison for Classifying Peatland Vegetation Types Using In Situ Hyperspectral Measurements”. In: Remote Sensing 9.7, p. 748.

  Harris, A., R. Charnock, and R. Lucas (2015). “Hyperspectral remote sensing of peatland floristic gradients”. In: Remote Sensing of Environment 162, pp. 99–111.

  Schmidtlein, S., H. Feilhauer, and H. Bruelheide (2012). “Mapping plant strategy types

using remote sensing”. In: Journal of Vegetation Science 23.3, pp. 395–405.

They have studied link between spectral response and species diversity. Considering peatland, identification of unique species using remote sensing does not make sense. That's why different approach such as mapping continuum or using different machine learning algorithm to produce probabilistic map are investigated.

6.3. Hyperspectral characterisation and mapping of plant functional types

l475: it's frustrating to read "the variations in spectral response" without ever seeing them illustrated…

7. References

Remove double references.

Check doi and complete where missing.

Comments for author File: Comments.pdf

Author Response

Author response to Reviewer 1 comments, 05/06/19

We would like to thank the reviewer for their feedback on our manuscript. We have made every effort to address their comments and criticisms to the best of our ability. In particular, we thank the reviewer for altering us to the presence of additional studies similar to ours. We have incorporated those suggestions into our introduction and discussion sections. We have also provided several additional plots in supplemental materials to address the comments made by multiple reviewers.

The reviewer suggests providing more details on the vegetation and diversity differences among treatments, particularly at the SPRUCE experiment. While we acknowledge that this would be a valuable addition to this study, we currently have a manuscript in preparation that details the vegetation changes with warming treatment at SPRUCE. Therefore, we would prefer not to publish those data in this manuscript.

Responses by M. McPartland are provided after each comment. Line numbers referenced in responses reflect line numbers with track changes enabled in MS Word.

1. General impression

The paper puts into perspective topics that will be increasingly highlighted because of environmental problems. What can be done to link ecological problems with remote sensing techniques that allow for long-term monitoring? (EBV's should be mentioned) The difficulty of the subject and the challenge raised must be highlighted: peat bogs have highly heterogeneous plant communities and the spatial scales used in hyperspectral remote sensing are not necessarily adapted.

M. McPartland - We have added discussion on the importance of peatland heterogeneity (68-71), as well as the use of hyperspectral methods in developing essential biodiversity variables (135-138).

The manuscript should be published after some modifications. Link between spectral signatures and ecological properties/diversity would be appreciated. Moreover, there is a lack of data on the percentage of cover or the different Shannon diversities for example. However this study show encouraging results.

M. McPartland – We are reluctant to publish details on the vegetation cover changes because we are currently preparing a manuscript that details this information for publication.

2. Abstract

l34-37: can the authors quantify ``strong effect'' and ``positive relationship'' ?

M. McPartland – We are unsure as to whether it is appropriate to add model results to an abstract. Since we have two experiments and several analyses that we performed, that would require us to add quite a number of model results to the abstract. We will do this if the editors decide that it is necessary.

3. Introduction

A detailed plan at the end of the introduction would be appreciated.

M. McPartland. We have edited the final paragraph of the introduction to include a more detailed plan including analysis details (139-155)

4. Methods

4.1. Study Sites

A first map to locate the sites would help. The presentation of the data is a little confusing:

M. McPartland – In an effort to address the comments by multiple reviewers regarding our methods description, we have added a map that shows the location of both sites (Figure 1).

APEX consists of 3 water manipulation: it is not clear that these treatments are studied or used later in the article.

              M. McPartland – Language edited for clarity (166-167)

SPRUCE: in total, ``we tracked 33 vegetation community plots…'' if this is the case, it should be mentioned in the following section.

              M. McPartland – Line removed, and specific reference to number of plots tracked provided (228).

4.2. Data Collection

4.2.1. Vegetation Cover Sampling

In addition to Table 1, a summary table of the different used data would provide a better understanding: e.g. APEX, number of plots per treatment, percent cover,…

M. McPartland – We have added an additional table (Table 1) that lists the experiments and data collection methods at each.

l188: how many plots ? map to locate transects ?

M. McPartland – We have deleted this line for clarity. Information on plots sampled at SPRUCE provided (226-233). In addition, we have added images of the sites in an effort to help contextualize the experiments (Figure 2).

It is regrettable not to have a table of floristic records for SPRUCE. Could it be added as supplementary materials ?

M. McPartland – We are reluctant to provide details of the vegetation change at SPRUCE because these results are currently in preparation in a manuscript to be submitted to an ecology journal.

As for APEX, a map to locate the plots would be appreciated.

M. McPartland – We have added Figure 1, 2, and Table 1 to provide additional context and aid in the interpretation of the experiments.

Should it be possible to have a summary table with the dominant species or the PFT depending on the plots for the 2 study sites?

              M. McPartland – We provide Table 2 with a species list.

4.2.2. Spectral Sampling

Could the authors give instrumental characteristics of the spectroradiometer (spectral resolution and spectral sampling according to the spectral range)?

M. McPartland – We have opted to restructure this section for clarity, following the comments made my multiple reviewers. We provide instrumental details on lines 240-241 and 274-276.

figure1: Wavelength instead of Bands

              M. McPartland – change made (Figure 3)

l218: \SI{44}{\centi\metre} Eq. (1): \[2 \times \textrm{height} \times \textrm{tan}\left( \dfrac{25\si{\degree}}{2} \right)\]

M. McPartland – We have deleted Equations 1 and 2 following the guidance of Reviewer 4. We have opted instead to cite a previous paper in which these data were also used (McPartland et al. 2018, Global Change Biology).

How many samples were acquired during one measurement?

              M. McPartland – Three scans were taken per measurement. Details provided at line 244 and 276.

l226: Did the three measurements performed on the same plot or ``around'' to capture vegetation variability?

              M. McPartland – Three scans were taken per measurement. Details provided at line 244 and 276.

l224: Did the measurements take place over several days? If so, it should be mentioned.

M. McPartland –  The measurements took place on one day. We have attempted to clarify this detail by providing the specific dates of data collection at lines 241 and 270.

l227-228: did the authors use an atmospheric transmittance model to remove the bands ? if so, which one ?

M. McPartland – We did not use an atmospheric transmittance model. We simply excluded regions that presented with a lot of noise associated with the instrument. Method described at line 250-254.

l233: As for the previous instrument (ASD), could the authors provide instrumental specifications ?

              M. McPartland – Detail provided at line 274-276.

As the acquisition dates differ and the sites are not at the same latitude and longitude, it should be ensured that the position of the sun does not have an impact on the measurements. The authors should specify the position of the Sun (azimuth angle).

              M. McPartland – Information added at line 245 and line 272.

l255: Did the measurements take place over several days? If so, it should be mentioned.

M. McPartland –  The measurements took place on one day. We have attempted to clarify this detail by providing the specific dates of data collection at lines 241 and 270.

4.2.3. Aerial hyperspectral data collection

Did the hyperspectral data collected only on APEX site? Did the imagery collected on June of 2015 or July of 2014 ?

              M. McPartland – Discrepancy resolved. Thanks for noticing.

l266: Could the authors provide the instrumental characteristics of the hyperspectral camera ? Is the camera composed of two captors ? In that case, spectral resolution and spectral sampling are not the same in the VNIR and the SWIR. Besides, are the authors sure of the altitude of the flight ? \SI{200}{\metre}a.g.l. ?

              M. McPartland – Additional information provided about the instrument (lines 309 – 313).

l272: could the authors precise the plugin ? Semi automatic classification ? dzetsaka ?

              M. McPartland – Method added at line 316-317.

4.3. Data analysis

4.3.1. Spectral data analysis

l281: analysis

Could the author add reference about constrained principal component analysis ?

              M. McPartland – References added at line 325

l290: is alpha-value different of p-value (which is usually used)?

              M. McPartland – Language clarified (line 332).

l302: ANOVA

              M. McPartland – Typo fixed

l312: are there enough points to confirm normal distribution ? I don't understand what is normally distributed: spectral coefficient of variations, spectral heterogeneity, community heterogeneity per PFT ? per species ? all data ? Could the authors provide the diagnostic plots ?

M. McPartland – we have provided diagnostic quantile-quantile plots in the supplemental materials to demonstrate that our data are distributed normally (Figure S4).

4.3.2. Hyperspectral image analysis

l321: could the authors convert into meters ?

              M. McPartland – Change made.

l323: which software was used to mosaic and georeference images ? which parameters were used for RF (number of estimators or trees, depth of each tree, …)?

              M. McPartland – Information on software added (line 364), and on RF parameters (369).

5. Results

5.1. Community drivers of spectral response

Could the authors show the different spectral signatures ?

M. McPartland – Figures added to supplemental materials to show plant spectra for all treatments and cover types (Figures S1, S2, S3).

The relationship between PCA and explained variables is not clear. Can the author detail how variables can be explained by reducing spectral dimension ?

              M. McPartland – Additional language added lines 389 – 390.

5.2. Relationships between species diversity and spectral variation

f_11 means F-score ? if so, authors should mention it or detail it.

M. McPartland – F values refer to the model degrees of freedom in the linear model. Language added for clarity line 448.

Can the "not difference" seen be explained by different location of sites and different dates of measurements and different "conditionment" ?

M. McPartland – Since our measurements were performed on the same day (language added in Methods for clarity), it couldn’t be different dates of measurement. We believe our lack of statistical significance stems from our relatively low sample size. No change were made in response to this comment because none was suggested.

5.3. Hyperspectral image analysis – mapping of PTFs

l405 - figure 8: how about ground truth to compare ?

              M. McPartland – Unfortunately, additional field work will not be possible on this project.

RF is sensitive to the ratio of sampling the individuals in population and parameters. This is why it is extremely interesting to specify the number of individuals as well as the parameters inherent to the algorithm.

              M. McPartland – We are unclear as to what suggestion the reviewer is making here.

6. Discussion

6.1. Hyperspectral remote sensing of peatland to experimentally-manipulated climate change

It is difficult to conclude because the sites do not benefit from the same acquisition protocols. Moreover, the number of individuals in the different samples does not seem sufficient to conclude (given the large number of hyperspectral parameters in terms of wavelengths).

M. McPartland – We are not particularly arguing that the sites should be directly comparable, since, as the reviewer states, they are subject to different data collection protocols. Although we didn’t have a very large sample size at APEX, at SPRUCE we had over 30 plots that were included in the analysis. At both sites, we found a strong effect of PFT cover on reflectance. Therefore, we believe that our results support our discussion,

6.2. Remote sensing of boreal peatland species diversity

Authors should either mentioned:

  Erudel, T., S. Fabre, T. Houet, F. Mazier, and X. Briottet (2017). “Criteria Comparison for Classifying Peatland Vegetation Types Using In Situ Hyperspectral Measurements”. In: Remote Sensing 9.7, p. 748.

  Harris, A., R. Charnock, and R. Lucas (2015). “Hyperspectral remote sensing of peatland floristic gradients”. In: Remote Sensing of Environment 162, pp. 99–111.

  Schmidtlein, S., H. Feilhauer, and H. Bruelheide (2012). “Mapping plant strategy types using remote sensing”. In: Journal of Vegetation Science 23.3, pp. 395–405.

              M. McPartland – references added lines 483-485 and 504.

They have studied link between spectral response and species diversity. Considering peatland, identification of unique species using remote sensing does not make sense. That's why different approach such as mapping continuum or using different machine learning algorithm to produce probabilistic map are investigated.

M. McPartland – Thank you for the suggestions. We are happy to see other studies that have taken a similar approach. We use the PFT framework as a more meaningful method of quantifying peatland vegetation changes than species identification.

6.3. Hyperspectral characterization and mapping of plant functional types

l475: it's frustrating to read "the variations in spectral response" without ever seeing them illustrated…

              M. McPartland – Vegetation type spectra provided in supplemental materials (Figure S1).

7. References

Remove double references.

              M. McPartland – Changes made

Check doi and complete where missing.

              M. McPartland – DOIs added to all references. 

Author Response File: Author Response.docx

Reviewer 2 Report

Major comments: 

At the methodology, you need to better explain how the training sample was taken; is it only by visual evaluation of the drone images? if yes, how confident you are that the interpretation was realistic? 

The dataset presented and the methodology do not explain how the validation of the supervised classification was performed. This is an important part of every classification.

On the classification of the hyperspectral image, why was the whole spectrum considered for the analysis? In the field hyperspectral data, the water-absorption bands and noise were disregarded. Shouldn't this be the case for the airborne data as well? 

There are several dates of data collection, which apparently associate different environmental and ecosystem conditions. The effect of the time difference and how and if this is related to erroneous interpretation should be discussed. 

Minor comments:

line 58 "and livelihoods and local communities" -> "and livelihoods of local communities"

line 84: what type of "High-resolution" are you referring to, spatial, radiometric or temporal? and how has this been emerging lately? I assume you do not refer to spectral, as the high spectral resolution is an intrinsic characteristic of a hyper-spectral system. 

line 107: "a number approaches to" -> "a number of approaches to"

line 215: delete "that reflected 100 percent of incoming solar radiation", it is implied by the white spectralon panel; moreover, no natural or artificial surface can reflect 100% from a physical standpoint of view. 

line 217: it doesn't really matter how many meters above the ground the measurement was taken; what matters is how many meters above the top layer of the vegetation the measurement was taken, since the wavelengths you are looking (visible and NIR) have no penetration to reach the ground surface. 

line 220: at the x-axis of the figure, replace "Band" with "Wavelength". Also the number of points are minimal for hyperspectral data; I would expect that an ASD with 1nm spectral bandwidth would have given hundreds of data points in the atmospheric window you are looking. Has a spectral re-sampling been carried out? 

line 265 refers to 2015 as a date of data collection, while line 269 to 2014; 

line 276: would help to integrate a larger map of Alaska to show where exactly this place is, given the fact that most readers are not very familiar with Alaska...

line 302: perhaps ANOVA instead of ANVOA?

line 324: you define "random forests (RF)" at line 315 earlier. 

line 474: perhaps instead of "near-surface" replace with "close-range. "

Author Response

Author response to Reviewer 2 comments, 05/06/19

We appreciate the reviewer for their feedback on our manuscript. We have made every effort to address criticisms to the best of our ability. The reviewer points out correctly that visually selecting the training points from the UAV image was likely to be less accurate than point selection in the field. This is a valid criticism of our methodology. Unfortunately, we are unable to provide a better training sample for our supervised classification analysis. We initially performed an unsupervised classification analysis on the hyperspectral dataset using both a subset of bands selected through a dimensionality-reduction analysis, and with the full stack of spectral bands. We found that we were able to substantially improve the quality of the final model result by incorporating the training pixels. Although there remains some uncertainty regarding the quality of those pixels, we believe that our results are still valid. As was pointed out by Reviewer #2, given the heterogeneous nature of peatland complexes, it is certain that there are compositionally mixed pixels in the aerial dataset as well. Given the constraints on our research, as well as the time delay between the time of data collection and today, we are unable to provide a ground-truthed training sample for our supervised classification analysis. We hope that this explanation will be sufficient to the reviewer and editor.

Responses by M. McPartland are provided after each comment. Line numbers referenced in responses reflect line numbers with track changes enabled in MS Word.

At the methodology, you need to better explain how the training sample was taken; is it only by visual evaluation of the drone images? if yes, how confident you are that the interpretation was realistic?

M. McPartland – More explanation added at line 360-365. See above explanation in response to comment.

The dataset presented and the methodology do not explain how the validation of the supervised classification was performed. This is an important part of every classification.

              M. McPartland – Details on model performance included (line 371-377).  

On the classification of the hyperspectral image, why was the whole spectrum considered for the analysis? In the field hyperspectral data, the water-absorption bands and noise were disregarded. Shouldn't this be the case for the airborne data as well?

M. McPartland - In regards to the exclusion of certain regions of plot-level but not the aerial hyperspectral dataset, we processed our datasets to reduce noise introduced either by atmospheric interference or by the instrument. This was an issue with the ASD, but was not an issue with the ProSpecTir imaging spectrometer. Post-processing of the aerial hyperspectral data also involved an atmospheric correction model (details provided in Anderson et al. in review at Remote Sensing of Environment). We feel that it was appropriate to process the data in two different ways, and find that including those regions in the final land cover classification is justified.

There are several dates of data collection, which apparently associate different environmental and ecosystem conditions. The effect of the time difference and how and if this is related to erroneous interpretation should be discussed.

M. McPartland – In response to comments by multiple reviewers, we have edited the methods sections to be more explicit regarding the dates of data collection. In general, we sought to source data collected at peak growing season. The only exception is in the case of the SPRUCE hyperspectral dataset, where data were collected in September. In this case, we found that the spectral distinction among cover types was in fact more pronounced, and posit in our discussion that this might be a feature of collecting hyperspectral data late in the season when there is optical differentiation associated with leaf senescence. 

Minor comments:

line 58 "and livelihoods and local communities" -> "and livelihoods of local communities"

              M. McPartland – Edit made (line 59).

line 84: what type of "High-resolution" are you referring to, spatial, radiometric or temporal? and how has this been emerging lately? I assume you do not refer to spectral, as the high spectral resolution is an intrinsic characteristic of a hyper-spectral system.

              M. McPartland – We were referring to high spatial resolution. Edit made for clarity (line 87).

line 107: "a number approaches to" -> "a number of approaches to"

              M. McPartland – Edit made (line 110)

line 215: delete "that reflected 100 percent of incoming solar radiation", it is implied by the white spectralon panel; moreover, no natural or artificial surface can reflect 100% from a physical standpoint of view.

              M. McPartland – Edit made (line 246)

line 217: it doesn't really matter how many meters above the ground the measurement was taken; what matters is how many meters above the top layer of the vegetation the measurement was taken, since the wavelengths you are looking (visible and NIR) have no penetration to reach the ground surface.

M. McPartland – Section edited and phrase removed in response to comments from multiple reviewers (lines 244-250).

line 220: at the x-axis of the figure, replace "Band" with "Wavelength". Also the number of points are minimal for hyperspectral data; I would expect that an ASD with 1nm spectral bandwidth would have given hundreds of data points in the atmospheric window you are looking. Has a spectral re-sampling been carried out?

M. McPartland – Axis label corrected (Figure 3). I also chose to make a line graph rather than a scatter plot to demonstrate the wavelengths included in the analysis. No spectral re-sampling as been carried out in this analysis.

line 265 refers to 2015 as a date of data collection, while line 269 to 2014;

              M. McPartland – Discrepancy resolved (line 302)

line 276: would help to integrate a larger map of Alaska to show where exactly this place is, given the fact that most readers are not very familiar with Alaska...

              M. McPartland – Figure 1 added to help orient readers to the location of both study sites.

line 302: perhaps ANOVA instead of ANVOA?

              M. McPartland – Typo fixed (line 343)

line 324: you define "random forests (RF)" at line 315 earlier.

              M. McPartland – Line removed (line 365)

line 474: perhaps instead of "near-surface" replace with "close-range. "

              M. McPartland – Change made (line 532).

Reviewer 3 Report

The manuscript by McPartland et al focuses on carrying out a study to map plant functional types in Boreal and Arctic vegetation using field spectroscopy and airborne hyperspectral data. The study is important due to its usefulness of mapping plant functional types of these changing ecosystems to a much larger area using airborne and/or spaceborne remote sensing. Caution has to be taken though, not to generalize algorithm developed in one area and apply across ecosystems. Heterogeneity of ecosystems in terms of species composition, surface water presence and plant physiology takes a major role in driving these empirical relationships developed at plot scale.

The authors do an excellent job in presenting the study and providing details of the methodology and results. The discussion of the manuscript is also well-written. The data and analysis are well presented and hence makes it a study well worth publishing without any major modifications.

I would like to congratulate the authors in carrying out an important work and presenting thoroughly to the readers.

Author Response

Author response to reviewer comments, 05/06/19

We thank the reviewer for their praise of our manuscript. We certainly agree with their main comment of the importance of accounting for landscape heterogeneity when scaling between plot-level and aerial datasets. In response to this and other reviewer comments, we have edited our introduction to include a more thorough discussion of heterogeneity in peatland systems (line 67-70).

The manuscript by McPartland et al focuses on carrying out a study to map plant functional types in Boreal and Arctic vegetation using field spectroscopy and airborne hyperspectral data. The study is important due to its usefulness of mapping plant functional types of these changing ecosystems to a much larger area using airborne and/or spaceborne remote sensing. Caution has to be taken though, not to generalize algorithm developed in one area and apply across ecosystems. Heterogeneity of ecosystems in terms of species composition, surface water presence and plant physiology takes a major role in driving these empirical relationships developed at plot scale.

The authors do an excellent job in presenting the study and providing details of the methodology and results. The discussion of the manuscript is also well-written. The data and analysis are well presented and hence makes it a study well worth publishing without any major modifications.

I would like to congratulate the authors in carrying out an important work and presenting thoroughly to the readers.

Reviewer 4 Report

McPartland et al. studied the effect of climate change in peatland reflectance. They used two boreal experiments to assess the reflectance characteristics over a gradient of species diversity and plant functional characteristics. The paper is generally well written, while the applied methods are meaningful to answer the research questions. Nevertheless, the manuscript is too long at some sections,  where there is still some level of repetition. I would reduce the number of figures as well; for example, Figure 2, 3, and 4 could be merge, and also Figure 5 and 6.

I also found that the PFT classification analysis (Figure 8) fell a bit off-topic regarding the research questions, so I suggest to drop it as is, in my opinion, resting strength to the manuscript. This analysis shows that the identification of discrete vegetation classes with hyperspectral data is possible, which is not per se novel. Instead, the authors could extrapolate the specific finding of the PCA and diversity experiments by regression analyses of the ordination axis. This would allow to further discuss the gradient links between reflectance and PFT-traits-diversity found during the analyses. For example, refer to:

Feilhauer, H., Faude, U., Schmidtlein, S., 2011. Combining Isomap ordination and imaging spectroscopy to map continuous floristic gradients in a heterogeneous landscape. Remote Sens. Environ. 115, 2513–2524. https://doi.org/10.1016/j.rse.2011.05.011

Schmidtlein, S., Feilhauer, H., Bruelheide, H., 2012. Mapping plant strategy types using remote sensing. J. Veg. Sci. 23, 395–405. https://doi.org/10.1111/j.1654-1103.2011.01370.x

Specific comments:

L30-34: These two sentences could be merge into one.

2.2.2. Spectral sampling: Equations 1 and 2 could be obviate as is a rather common procedure and are not used later on in the manuscript. The authors can refer to previous papers of McPartland (e.g. McPartland et al., 2018; GCB) for specific methods and equations, so this section could be reduced.

262-264: Information on floristic composition was already presented.

268: This is confusing, what do you mean by 5 nm spectral resolution? You said that the bandwidth was 1 nm already. Spectral resolution refers to the bandwidth, the number of bands, and the spectral cover altogether. Please clarify.

318-320: Did you consider the the size of the objects when locating the training points with the UAV? Considering that the hyperspectral flight has 1 m pixels of size, there is the chance of mixed class pixels.

384-393: This belong to methods.

395: Significant levels do not have a gradient. Hence, a relationship is or is not significant according to a certain criteria. Delete ‘if only marginally significant’.

405: Figure 8 does not show how successful was the classification.

468-470: This is a bold statement. I would soften this as there are actually many diversity studies in non-woody environments.

477-478: I don't see the logical link here.

481-483: Revise content of this phrase, merge with previous one.

487-488: Whats the meaning of this? By representative you mean that there is a high agreement (e.g. Kappa) between the observed and predicted PFT covers?

Author Response

Author response to Reviewer 4 comments, 05/06/19

We appreciate the reviewer for their feedback on our manuscript. We have made every effort to address the comments and criticisms to the best of our ability. In particular, we thank the reviewer for alerting us to the presence of studies similar to ours. We have incorporated those suggestions into our framing and discussion. We have taken the reviewer’s suggestion to decrease the size and number of figures. We have also tried to streamline certain sections, in particular the methods section, to increase the flow and quality. In lieu of an exhaustive methods description, we have instead referenced a previous paper published that describes the same data collection procedures. We have also tried to decrease repetition throughout the writing.

The reviewer suggested that we remove the final section of our analysis, in which we used an aerial dataset to map PFT distribution across a landscape, and do further analysis of our plot-level hyperspectral data. We have opted to maintain this section, but have added additional discussion and citations as the reviewer suggests. We hope that we will have sufficiently addressed the reviewer’s suggestions, without having to exclude our aerial hyperspectral data from the manuscript.

Responses by M. McPartland are provided after each comment. Line numbers referenced in responses reflect line numbers with track changes enabled in MS Word.

McPartland et al. studied the effect of climate change in peatland reflectance. They used two boreal experiments to assess the reflectance characteristics over a gradient of species diversity and plant functional characteristics. The paper is generally well written, while the applied methods are meaningful to answer the research questions. Nevertheless, the manuscript is too long at some sections, where there is still some level of repetition. I would reduce the number of figures as well; for example, Figure 2, 3, and 4 could be merge, and also Figure 5 and 6.

M. McPartland – We have merged a number of the figures by creating side-by-side panels instead of full-page figures. We have also tried to streamline the language in the introduction and discussion to avoid repetition.

I also found that the PFT classification analysis (Figure 8) fell a bit off-topic regarding the research questions, so I suggest to drop it as is, in my opinion, resting strength to the manuscript. This analysis shows that the identification of discrete vegetation classes with hyperspectral data is possible, which is not per se novel. Instead, the authors could extrapolate the specific finding of the PCA and diversity experiments by regression analyses of the ordination axis. This would allow to further discuss the gradient links between reflectance and PFT-traits-diversity found during the analyses. For example, refer to:

Feilhauer, H., Faude, U., Schmidtlein, S., 2011. Combining Isomap ordination and imaging spectroscopy to map continuous floristic gradients in a heterogeneous landscape. Remote Sens. Environ. 115, 2513–2524. https://doi.org/10.1016/j.rse.2011.05.011

Schmidtlein, S., Feilhauer, H., Bruelheide, H., 2012. Mapping plant strategy types using remote sensing. J. Veg. Sci. 23, 395–405. https://doi.org/10.1111/j.1654-1103.2011.01370.x

M. McPartland – We have incorporated these citations and associated discussion of references’ findings into our discussion section (lines XXX-XXX). We have opted not to drop the final section of our analysis. While we know that this approach is not novel, as are other parts of this study, we still find that it complements our plot-level research to demonstrate that PFT classification is scalable using aerial datasets. Furthermore, none of the other reviewers suggested to remove that section. We hope that the editors will find our decision acceptable.

Specific comments:

L30-34: These two sentences could be merge into one.

              M. McPartland – Change made (lines 32-34).

2.2.2. Spectral sampling: Equations 1 and 2 could be obviate as is a rather common procedure and are not used later on in the manuscript. The authors can refer to previous papers of McPartland (e.g. McPartland et al., 2018; GCB) for specific methods and equations, so this section could be reduced.

M. McPartland – We have significantly re-worked this entire section following the reviewer’s suggestion. We have removed some descriptions of our hyperspectral data collection and processing methods and instead provided citations to the papers suggested.

262-264: Information on floristic composition was already presented.

              M. McPartland – Text edited (line 305).

268: This is confusing, what do you mean by 5 nm spectral resolution? You said that the bandwidth was 1 nm already. Spectral resolution refers to the bandwidth, the number of bands, and the spectral cover altogether. Please clarify.

              M. McPartland – Language edited for clarity

318-320: Did you consider the the size of the objects when locating the training points with the UAV? Considering that the hyperspectral flight has 1 m pixels of size, there is the chance of mixed class pixels.

M. McPartland – It is certainly possible that there might have been mixed pixels, given the differences in resolution between the two datasets. This is a constraint of the system, since there are very unlikely to be “pure” pixels given the floristic gradients present across the space.

384-393: This belong to methods.

M. McPartland – Phrases deleted to avoid repetition. Methods described lines 333-353.

395: Significant levels do not have a gradient. Hence, a relationship is or is not significant according to a certain criteria. Delete ‘if only marginally significant’.

              M. McPartland – Language edited (line 447).

405: Figure 8 does not show how successful was the classification.

              M. McPartland – Classification results are presented in Table 5 and 6.

468-470: This is a bold statement. I would soften this as there are actually many diversity studies in non-woody environments.

              M. McPartland – Section edited (lines 522).

477-478: I don't see the logical link here.

              M. McPartland – language edited (line 536).

481-483: Revise content of this phrase, merge with previous one.

              M. McPartland – language edited (line 536-539).

487-488: What’s the meaning of this? By representative you mean that there is a high agreement (e.g. Kappa) between the observed and predicted PFT covers?

              M. McPartland – Language edited for clarity (line 546).

Round 2

Reviewer 1 Report

Reviewer response to authors response, 05/13/19.

\newline{}

The authors' efforts to correct and take into account the comments are appreciated. However, there are still inaccuracies or a lack of relevant data to support some of the statements. They are listed below. 

 - the fact that the authors are reluctant to publish vegetation data in a comprehensive way because another article is in preparation can be understood. However, it seems crucial to me to have a minimum indication (percentage of majority species, species diversity for a given plot) in order to better understand the link that may exist with spectral signatures.

 - Eq. (1): \citep{mcpartland2019response} contains a typographical error. I suggest to insert the right equation in the present article: \[2 \times \textrm{height} \times \textrm{tan}\left( \dfrac{25\si{\degree}}{2} \right).\] 

 - \citep{mcpartland2019response} details the data acquisition protocol using the spectroradiometer (``Three scans were performed above each vegetation plot and averaged``) but they do not specify whether the 3 measurements were made at  or around the same location point (in this case over what distance?) to measure the spectral heterogeneity due to floristic heterogeneity (what would then be approximately the extent covered ?). If it is correctly understood, it is not a spatial average but a local average (whereas the instrument already averages several acquisitions during a single measurement). 

 - detailed information on the geolocation and post-processing of the aerial hyperspectral data is provided in a futur article. But is it a submitted and accepted article subject to (major) modification or only a submitted article? Besides according to the data provided by the authors and the camera supplier, a pixel of \SI{1}{\metre^2} does not seem feasible.

 - the use of Spectral Angle Mapper to post-process data is not very clear because Spectral Angle Mapper is rather a tool used to classify pixels or select pixels belonging to the same class. Moreover, the complete reference of the plugin or link would be appreciated as it does not match any QGIS 3 plugin. 

 - the article referring to ``constrained PCA`` is still missing. The authors mention two types of dimension reduction methods \citep{harsanyi1994hyperspectral,li2012locality} but it is not clear what the constraint for PCA is? why not do a simple PCA if not? Moreover, the results of the graphs with the PCA axes are not very explicit.

 - when the sample size is very small, it is advisable to use appropriate statistical tests (such as Anderson-Darling for example). To use a QQ-plot, which is a visualization tool, you need at least 30 points.

 - the part relating to Random Forest (RF) should be removed, as it seems that validation measurements have been made on the classifier's training data set (the authors mentioned that they selected about 100 points on the image; according to Table 5, about 100 points per class are used to validate). Moreover, in the absence of ground truth points (taken either in the field or by photo-interpretation), it is difficult to draw conclusions. l327: Square root RF of the number of parameters rather.

l67: Plant functional types are first mentionned here, so acronym should be here

not l68 ; typo. : Erudel \emph{et al.} 2015

Comments for author File: Comments.pdf

Reviewer 2 Report

Authors have addressed the issues raised. It is understood that the training dataset was collected a postieriori and cannot use training dataset of simultaneous acquisition, although this would considerably boost the scientific soundness of this study. I feel line the sentence introduced 'Using the aerial image also allowed us to identify a far larger number of training points per class than would have been feasible in the field.' should be deleted; stating that the number of training pixels is higher when preferring visual evaluation rather than actual fieldwork as the ground truth data collection, is not a very valid point; a surveyor could collect as many point pixels as necessary for a classification task. 

Reviewer 4 Report

McPartland et al. improved their manuscript in terms of structure and clarity. I still have some questions regarding results (APEX ordination results, see specific comments) and the analysis could be improved at some points (see also specific comments). I also found some false citations in the discussion section, just because I am familiar with the cited work. Please recheck the citations throughout the manuscript. Bad citations could undermine your work after publication.

Specific comments:

L193: Remove ‘somewhat’.

Table 2: ‘List of vascular plant species…’

L351: You stated here that the spectral variation at APEX is related to the cover of mosses and forbs, but forbs seems not to be highly correlated to reflectance according to the vectors length (Figure 3) and the permutation test. Instead, just mosses and litter seems to have the largest explained variance. Forbs, Equisetum, Sedges and Shrubs are so close in the feature space that making further conclusions regarding these PFTs seems, in my opinion, misleading. At this level, the smaller differences are probably caused by site-specific conditions, hence I recommend to center only in the significant and the

apparent tendencies. For example, as Litter was included in the analysis, it is expected that the main reflectance gradient would differentiate mainly between alive-dead (green-nongreen) vegetation. You could perform a second PCA dropping the Litter values to enhance the PFT analysis.

L426-430: Related to previous comment; it is clear that the main PCA gradient will show GPP and cover gradients when non-vegetated areas are included in the analysis (here Litter).

L430-431: This is clear, as SPRUCE is a rich poor environment.

L431-432: Why are shrubs and forbs explaining the gradient explaining the reflectance shifts under climate change conditions? Are these PFTs more competitive under high temperatures and lower water tables? Is it because the change on soil microbial activity? Please go deeper in the interpretations of the found links.

L443: Does this study really demonstrate that drought in peatlands negatively affects diversity? I cannot see the results supporting this statement directly. This conclusion could be obtain by a direct analysis of the species pool across the experiment treatments, but the comprehensive list is not presented here. Please constrain your conclusions to the presented results, or provide a better link to understand how such conclusions were drew from the data.

L452-453: Are there other studies with similar approaches?

L462-463: Neither of these three citations created estimated richness directly or produces probabilistic maps of vegetation patterns. Both Feihauer et al. 2011 and Harris et al. 2015 mapped the floristic composition of peatlands by using the axis of ordinations as response variables and reflectance as predictors. Therefore, predicted values are scores of floristic composition in the feature space (i.e. species presences). Meanwhile, Erudel et al. 2017 tested different combinations of spectral distances, transformations and classifiers to classify PFTs in peatlands. While some of the algorithms used probabilistic assumption in their internal process, a final map of probability was not presented. Please be accurate when citing relevant previous works, as this could affect the performance of this paper in the future.

L466: Which are the different spatial scales? I would argue that the whole analysis was performed at the same spatial scale (1-2 m).

L469-472: What about the effect of structural traits on PFT discrimination? Recent works pointed out that structural traits may have grated effects on reflectance than leaf traits to separate between PFTs:

Kattenborn, T., Fassnacht, F.E., Schmidtlein, S., 2018. Differentiating plant functional types using reflectance: which traits make the difference? Remote Sens. Ecol. Conserv. 1–15. https://doi.org/10.1002/rse2.86

477-479: Why do you think that shrubs were the most stable class to map? And how does this connect to the importance of shrubs in the spectral variation of SPRUCE? Discuss deeper in the interesting found connections.