Missing the Forest and the Trees: Utility, Limits and Caveats for Drone Imaging of Coastal Marine Ecosystems

Round 1

Reviewer 1 Report

The manuscript presents a good overview of pros and cons of the use of drones in coastal ecology.

Ecosystem managing is an extremely relevant topic, especially applied to coastal marine ecosystems, which are among the most endangered by both climate change and human pressure. In this context, drones imaging is gaining a primary role among the other remote sensing techniques, permitting high-resolution, frequent, and timely observations of large and relatively inaccessible areas.

This manuscript is considered a significant contribution to the journal, fully in line with its scope, it is very well written and easy to follow.

My only concerns are:

• I’m not an ecologist and I’m not familiar with the Bray-Curtis index or the PERMANOVA package. As such, I missed a more detailed description of the data analysis. My suggestion is add a specific paragraph including the definition of the Bray-Curtis resemblance matrix, and more details about the permutational analysis of variance, answering to: What’s PERMANOVA? How you define the pseudo-F? what does “6,39” stand for? Why pseudo-F=21 is considered a significant correlation?
• The description of colors and lines in fig. 3 is missing.
• What are the axis in Fig 4? 1^st and 2^nd Principal Component?

Author Response

Reviewer 1

The manuscript presents a good overview of pros and cons of the use of drones in coastal ecology.

Ecosystem managing is an extremely relevant topic, especially applied to coastal marine ecosystems, which are among the most endangered by both climate change and human pressure. In this context, drones imaging is gaining a primary role among the other remote sensing techniques, permitting high-resolution, frequent, and timely observations of large and relatively inaccessible areas.

This manuscript is considered a significant contribution to the journal, fully in line with its scope, it is very well written and easy to follow.

My only concerns are:

• I’m not an ecologist and I’m not familiar with the Bray-Curtis index or the PERMANOVA package. As such, I missed a more detailed description of the data analysis. My suggestion is add a specific paragraph including the definition of the Bray-Curtis resemblance matrix, and more details about the permutational analysis of variance, answering to: What’s PERMANOVA? How you define the pseudo-F? what does “6,39” stand for? Why pseudo-F=21 is considered a significant correlation?

These points have been clarified in the methods (see lines 233,238)

• The description of colors and lines in fig. 3 is missing.

Done, thanks

• What are the axis in Fig 4? 1^stand 2^nd Principal Component?

This has been better described in the figure legend

Reviewer 2 Report

• Please add a more thorough explanation to lines 44-54 about why the bull kelp species matters explaining why it is the primary classification target of this study. This reviewer did not find adequate explanation of WHY this species matters both from an ecological perspective and from a monitoring/remote classification perspective. Does this species have a distinct spectral signature that makes it advantageous for this study? Is this species especially vulnerable to climate change, or it is valuable for mitigation of the effects of climate change?
• The discussion is nicely structured with an elegant presentation of the gains, trade-offs, and limitations of drone/UAV imagery for this type of classification analysis. However, there is repeated emphasis throughout the discussion on the limitations of this technique to measure biodiversity and conduct a full taxonomic inventory (lines 286-288, 299-301, and 303). This is then repeated as the 3rd bulleted item under the limitations portion of the discussion (lines 316-317). Please consider only mentioning this concept once as a limitation. Or provide equal amounts of discussion for all three limitations.
• In lines 344 and 345, the authors mention that the limitations and caveats of drone use for this type of monitoring and survey analysis will decrease over time; however, no evidence is presented for why the authors believe this trend to be true. Please revise wording or provide a discussion of the contributing factors to justify the decrease over time of the limitations and trade-offs presented. Why will the significance of your findings decrease in time?

Author Response

Reviewer 2

  Please add a more thorough explanation to lines 44-54 about why the bull kelp species matters explaining why it is the primary classification target of this study. This reviewer did not find adequate explanation of WHY this species matters both from an ecological perspective and from a monitoring/remote classification perspective. Does this species have a distinct spectral signature that makes it advantageous for this study? Is this species especially vulnerable to climate change, or it is valuable for mitigation of the effects of climate change?

 

The original manuscript has outlined the massive loss of this species due to both heat wave events, and earthquake uplift. These impacts are of such magnitude that entire populations have been lost, or severely reduced. We also point out the challenging nature of the locations that this species inhabits (Figure 2) providing critical justification for why remote tools are appropriate. As per the reviewers comments we have further justified the important global ecological role of this species and other macroalgae and have moved the text to make these points earlier in the manuscript. See lines 39-55.

 

  The discussion is nicely structured with an elegant presentation of the gains, trade-offs, and limitations of drone/UAV imagery for this type of classification analysis. However, there is repeated emphasis throughout the discussion on the limitations of this technique to measure biodiversity and conduct a full taxonomic inventory (lines 286-288, 299-301, and 303). This is then repeated as the 3rd bulleted item under the limitations portion of the discussion (lines 316-317). Please consider only mentioning this concept once as a limitation. Or provide equal amounts of discussion for all three limitations.

 

We have consolidated this text to be less repetitive, but this is a key limitation not well covered by researchers elsewhere (see Lines 333-348). The resolution limitations and overlying water limitations are more obvious and documented and require little emphasis or explanation, whereas the complex structure of biogenic communities and the implications for using drone imaging for full taxonomic inventories is a key contribution that we believe can help guide management agencies and researchers when deciding on the appropriate tools to apply to coastal monitoring. 

 

  In lines 344 and 345, the authors mention that the limitations and caveats of drone use for this type of monitoring and survey analysis will decrease over time; however, no evidence is presented for why the authors believe this trend to be true. Please revise wording or provide a discussion of the contributing factors to justify the decrease over time of the limitations and trade-offs presented. Why will the significance of your findings decrease in time?

Yes, this is perhaps vague with little supporting evidence. We have modified this text with examples of why we think these tools will increase in utility through time. See lines 446-449.

Reviewer 3 Report

The authors employ drone imaging to classify the coastal seaweed species in several areas along New Zealand's west coast. They use RGB imaging and also multispectral imaging (6 bands plus NDVI) to train SVM classifiers using in situ sampling data. From those validated results, they measure biodiversity and test if it has been affected by the uplift caused by the 2016 earthquake. They show a few results (but link to the online ArcGIS map viewer which shows some spectacular results) and also show some correlations between uplift and cover percentage of different algal species. Their final contribution (and goal of the article) is the standardized methodology to obtain all this.

I believe the work has been well conducted and is worth being published. However, I would also like to recommend several improvements that I deem will make the article clearer for Remote Sensing readers.   Main questions:
First, the authors use drone images, however, most of the references cited in the introduction [12-16] refer to satellite imaging; no background on drone based coastal imaging is given in the introduction (although the authors have included some relevant references in the bibliography [17, 20, 21, 23, 24]). I think that would make it easier to appreciate the degree of innovation of the article methodology.   Second, the authors describe an imaging methodology based on RGB images which are photogrammetrically assembled to get 1.25 cm^2 pixel resolution images (I assume square pixel, correct?), as well as a digital elevation model (DEM) both computed with Agisoft Photoscan. They also describe the computation of stitched images (also a DEM?) using the multispectral images; afterwards they computed NDVI from reflectance normalized images. I do not see very clear what was the procedure from then on, and if the multispectral and the RGB images were finally used in all the analyses. Please, clarify this point. Subsection 3.1 seems to deal only with RGB images, until the "utility of multispectral imaging" is mentioned in L230 (should there be a subsection 3.2?). It is not clear whether the SVM classification procedure was the same for both types of images, or even if RGB and multispectral images were combined in some form (the first sentence of that paragraph could suggest that).   Third, the authors perform a multivariate analysis of the influence of earthquake uplift on the composition of habitat. I find this hypothesis missing some foundation. Without pre-2016 data to compare with, it is difficult to understand from the data in Figure 3 any causal relationship. I would probably ask if different coverages could not be more easily attributable to differences in intertidal coverage of those areas; those coverages would have been affected by the uplift, of course. But this hypotesis seems easier to test with the given data.   Other issues I suggest the authors to correct or check: L45: extra "is" (in "and is is an ecosystem dominant") should be deleted
L75: "develping" should be "developing"
L85-88: All the imaging sensors mentioned here are those of public agencies, whose images are provided open and free; however, there are other remote sensing private constellations that provide images with much higher spatial resolution (~30 cm). Some words about those sensors could be worth saying.
L96-28 (Figure 1): The caption (or the very image) should indicate that the lower-right images correspond to the dashed green rectangle on the upper-left RGB image.
L113: What is the CMOS sensor resolution (size of the images, ground pixel size)?
L126-128 (Table 1): In order to appreciate the importance of earthquake vertical deformation, please, include also information about the tidal amplitude in the study area (or, if there were significant variations between sites, add a fifth column to the table with that information)
L226-229 (Figure 4): Please, improve the description of the figure and the figure itself; what are the axes of the figure? L247: Are these all the results? I would appreciate some statistical summary, or some exploitation of the rich spatial information obtained from the classifications (e.g. neigboring relationships between species).
L346: Please cite some of the current marine spectral libraries already available.
L360-363: If this denotes the actual methodology used in this article, maybe this figure should go in that section. Otherwise, please note in the discussion what are the suggested improvements, or highlight there what other approaches were tested that were later discarded rendering this "distilled"

Author Response

Reviewer 3

The authors employ drone imaging to classify the coastal seaweed species in several areas along New Zealand's west coast. They use RGB imaging and also multispectral imaging (6 bands plus NDVI) to train SVM classifiers using in situ sampling data. From those validated results, they measure biodiversity and test if it has been affected by the uplift caused by the 2016 earthquake. They show a few results (but link to the online ArcGIS map viewer which shows some spectacular results) and also show some correlations between uplift and cover percentage of different algal species. Their final contribution (and goal of the article) is the standardized methodology to obtain all this.

I believe the work has been well conducted and is worth being published. However, I would also like to recommend several improvements that I deem will make the article clearer for Remote Sensing readers.   Main questions:
First, the authors use drone images, however, most of the references cited in the introduction [12-16] refer to satellite imaging; no background on drone based coastal imaging is given in the introduction (although the authors have included some relevant references in the bibliography [17, 20, 21, 23, 24]). I think that would make it easier to appreciate the degree of innovation of the article methodology.

Our point here is that where possible, satellites will provide the most cost-effective mechanism for monitoring ecosystems, but we outline why these methods aren’t always applicable for coastal ecosystems. We then outline why drones provide a suitable tool for collecting broad spatial scale information at an appropriate resolution. Perhaps this paragraph is somewhat laboured in making this point. We have condensed some of this text (see lines 83-87) to bring forward UAV-related research and citations (see lines 103-108).

 

   Second, the authors describe an imaging methodology based on RGB images which are photogrammetrically assembled to get 1.25 cm^2 pixel resolution images (I assume square pixel, correct?), as well as a digital elevation model (DEM) both computed with Agisoft Photoscan. They also describe the computation of stitched images (also a DEM?) using the multispectral images; afterwards they computed NDVI from reflectance normalized images. I do not see very clear what was the procedure from then on, and if the multispectral and the RGB images were finally used in all the analyses. Please, clarify this point. Subsection 3.1 seems to deal only with RGB images, until the "utility of multispectral imaging" is mentioned in L230 (should there be a subsection 3.2?). It is not clear whether the SVM classification procedure was the same for both types of images, or even if RGB and multispectral images were combined in some form (the first sentence of that paragraph could suggest that).

Thanks. We have better outlined the relative use of both RGB (see lines 159-165) and multispectral imagery (see lines 217-220) and where differences in analytical procedures have been used.

 

   Third, the authors perform a multivariate analysis of the influence of earthquake uplift on the composition of habitat. I find this hypothesis missing some foundation. Without pre-2016 data to compare with, it is difficult to understand from the data in Figure 3 any causal relationship. I would probably ask if different coverages could not be more easily attributable to differences in intertidal coverage of those areas; those coverages would have been affected by the uplift, of course. But this hypotesis seems easier to test with the given data.  

This is a good point and certainly more detail is required from our end. While we do not have pre-earthquake drone imagery, we have published pre- and post-earthquake survey data which reveals the dramatic loss of macroalgae, and the large increase in bare rock (Schiel et al. 2019). This manuscript uses our work along this earthquake-impacted coast as an example of the clear utility of these drone methods and the full context of the earthquake is difficult to describe without completely over-riding the manuscript. Brown macroalgae, such as the southern bull kelp, dominated the entire coastline (REFs) prior to the EQ, and the dramatic loss of this species was immediately obvious. What we show here is that the degree of vertical uplift had species (functional group) specific effects. We have further clarified these points to help the reader better interpret these results. See lines 119-124, lines 137-139

 

Other issues I suggest the authors to correct or check: 

L45: extra "is" (in "and is is an ecosystem dominant") should be deleted

Changed, thanks.

L75: "develping" should be "developing"

Changed, thanks.

L85-88: All the imaging sensors mentioned here are those of public agencies, whose images are provided open and free; however, there are other remote sensing private constellations that provide images with much higher spatial resolution (~30 cm). Some words about those sensors could be worth saying.

This is worth mentioning, and we have done so (Line 112, 424-427).

L96-28 (Figure 1): The caption (or the very image) should indicate that the lower-right images correspond to the dashed green rectangle on the upper-left RGB image.

Done

L113: What is the CMOS sensor resolution (size of the images, ground pixel size)?

Ground pixel size of c. 1 cm (see lines 156-157)

L126-128 (Table 1): In order to appreciate the importance of earthquake vertical deformation, please, include also information about the tidal amplitude in the study area (or, if there were significant variations between sites, add a fifth column to the table with that information)

The tidal amplitude is consistent along this whole coastline and the range of amplitudes has been added (See line 151).

L226-229 (Figure 4): Please, improve the description of the figure and the figure itself; what are the axes of the figure? 

The multidimensional scaling plot presents “coordinates” of similarity. This has been explained in the figure legend.

L247: Are these all the results? I would appreciate some statistical summary, or some exploitation of the rich spatial information obtained from the classifications (e.g. neigboring relationships between species).

We have added the classified output (Figure 5F), and have included accuracy estimates of the classification routines (Table 2).

L346: Please cite some of the current marine spectral libraries already available.

We have added multiple citations of studies building or using spectral libraries. See line 377.

L360-363: If this denotes the actual methodology used in this article, maybe this figure should go in that section. Otherwise, please note in the discussion what are the suggested improvements, or highlight there what other approaches were tested that were later discarded rendering this "distilled"

Yes this was the methodology implemented for this article, but it has taken some iterations to arrive at this point. In particular, we have used embedded transects in previous trials of these methods and found these to be limited to where we can place a transect without a wave moving it (see additional text lines 429-434). As the reviewer suggests, this could be easily moved to the methods. However, we believe that our methodology has utility for other researchers and that this workflow should be considered as a good starting point for others intending to use drones for conservation purposes. Part of the reasoning for placing this in the discussion is to promote this as a potential method which we believe leads to fit-for-purpose aerial imaging and robust accuracy assessment. We do, however, have no issue placing this in the methods if that is preferred.

Round 2

Reviewer 3 Report

I'm afraid the authors have not submitted the same version of the manuscript they refer to in their response, hence some of their responses have been hard to find as the lines do not match. 

The authors have addressed most of my previous concerns regarding the manuscript. They have clarified somewhat the motivation of using drone imaging, improved their bibliography revision and background on previous studies (UAV imaging, algae libraries), they have clarified a bit more their use of RGB and multispectral imaging, they have also largely improved their motivation of using multivariate analysis relating earthquake-caused uplift and habitat composition (their former studies of macroalgae loss in those areas after the earthquake), and provided some other information that can help the reader understand the research context. 

I believe the article is fine for publication. However, I would like to still suggest some improvements (order by relevance):

- The authors say they have better outlined the relative use of both RGB and multispectral imagery and the analytical procedures used. However, I still believe some clarifications are needed. In section 2.2 "Multispectral imaging", the authors describe the DJI Matrice 600 equipped with both a multispectral camera and an RGB camera. This RGB camera "was used to capture imagery", but is different from the Hasselblad CMOS RGB camera onboard the DJI Mavic Pro 2 drone. So when section "2.4. Analysis and Validation" (2.3?) starts with "RGB imagery (8 bit)", it is not clear whether it is the CMOS camera or the Sony RGB camera that is being used (or both). Perhaps the authors intended to make a distinction with the extra information of the images being 8 bit, but that information had not been given before. The input to the SVM, in terms of spectral bands should be better described: I guess that in the areas imaged from the DJI Matrice 600, they are fed in 3 RGB bands, 6 multispectral bands, and 1 NDVI calculated band, as a 10-band composite image (similar to ref. 23); then, in the areas imaged from the DJI Mavic Pro 2 drone, only the 3 RGB bands were used (?).

- The authors say that placing the methodology workflow in the discussion promotes it as a potential method that leads to fit-for-purpose aerial imaging and robust accuracy assessment. I understand that pointing that out is fine for the discussion (actually, it is a conclusion). However, I still believe that including the methodology workflow in the section 2 highlights that this is the actual workflow used successfully in the paper, not some suggestion for future improvements. So I again recommend that figure be moved to the end of materials and methods.

- Perhaps L121-127 is not the right place to describe the tidal regime of the study area or the weather and flight conditions. Section 2.1 "RGB imaging of rocky reef ecosystems" is dedicated to DJI Mavic Pro 2 drone RGB imaging, but I think all that environmental information is common to this and the next subsections; so my suggestion would be to move these details to somewhere around L108-109.

Author Response

I'm afraid the authors have not submitted the same version of the manuscript they refer to in their response, hence some of their responses have been hard to find as the lines do not match. 

Perhaps there has been confusion over which version of the manuscript that the line numbers referred to (there was a “tracked” version and a “clean” version), regardless we sincerely apologise for the frustrating inconvenience of searching through mismatched text. To be clear in our replies to these comments we are referring to line numbers in the “clean” version.

The authors have addressed most of my previous concerns regarding the manuscript. They have clarified somewhat the motivation of using drone imaging, improved their bibliography revision and background on previous studies (UAV imaging, algae libraries), they have clarified a bit more their use of RGB and multispectral imaging, they have also largely improved their motivation of using multivariate analysis relating earthquake-caused uplift and habitat composition (their former studies of macroalgae loss in those areas after the earthquake), and provided some other information that can help the reader understand the research context. 

I believe the article is fine for publication. However, I would like to still suggest some improvements (order by relevance):

- The authors say they have better outlined the relative use of both RGB and multispectral imagery and the analytical procedures used. However, I still believe some clarifications are needed. In section 2.2 "Multispectral imaging", the authors describe the DJI Matrice 600 equipped with both a multispectral camera and an RGB camera. This RGB camera "was used to capture imagery", but is different from the Hasselblad CMOS RGB camera onboard the DJI Mavic Pro 2 drone. So when section "2.4. Analysis and Validation" (2.3?) starts with "RGB imagery (8 bit)", it is not clear whether it is the CMOS camera or the Sony RGB camera that is being used (or both). Perhaps the authors intended to make a distinction with the extra information of the images being 8 bit, but that information had not been given before. The input to the SVM, in terms of spectral bands should be better described: I guess that in the areas imaged from the DJI Matrice 600, they are fed in 3 RGB bands, 6 multispectral bands, and 1 NDVI calculated band, as a 10-band composite image (similar to ref. 23); then, in the areas imaged from the DJI Mavic Pro 2 drone, only the 3 RGB bands were used (?).

The beginning of this paragraph, as you mention, wrongly emphasises the RGB imagery. In this section, the RGB imagery is used mainly as quality control (as mentioned Lines 190-191). We have now clarified this by starting the paragraph referring to the M600 drone flights with the dual camera system (Lines 172-173). We have also further clarified the multi-band products which were used as inputs to the SVMs for the RGB only analysis (Lines 134-136) and the multispectral product analysis (Lines 186-190).

- The authors say that placing the methodology workflow in the discussion promotes it as a potential method that leads to fit-for-purpose aerial imaging and robust accuracy assessment. I understand that pointing that out is fine for the discussion (actually, it is a conclusion). However, I still believe that including the methodology workflow in the section 2 highlights that this is the actual workflow used successfully in the paper, not some suggestion for future improvements. So I again recommend that figure be moved to the end of materials and methods.

We accept the reviewer’s assessment of the placement of this figure and have added it to the methods (See Lines 205-207).

- Perhaps L121-127 is not the right place to describe the tidal regime of the study area or the weather and flight conditions. Section 2.1 "RGB imaging of rocky reef ecosystems" is dedicated to DJI Mavic Pro 2 drone RGB imaging, but I think all that environmental information is common to this and the next subsections; so my suggestion would be to move these details to somewhere around L108-109.

This information is general across all drone deployments and has been moved to the general methods introductory section (see Lines 117-123).