Assessing the Potential Replacement of Laurel Forest by a Novel Ecosystem in the Steep Terrain of an Oceanic Island

Round 1

Reviewer 1 Report

I really appreciated the manuscript, which use the new approach of RS and field data to assess the current distribution and the potential spread of the non-native plant Castanea sativa in island of La Palma.

I only suggest to better explain the approach of RS data anaysis to people less confident with the use of RS. I'm not so familiar to give more precise suggestions. 

Please, check the autors names, there are some errors in the authors section and in the author contribution section.

Row 4-5  Please check the authors names.

Row 38-39 Please, delete the symbol of “~” from 1.2% and 4% and 13%

Row 58  Please, delete “species intrusion” that is not appropriated to the Invasive Alien Species terminology, “threatened by ecological invasion” is enough.

Row 72  Please, use the acronyms alone, by adding the list of acronyms at the beginning of the manuscript, or rather replace “Remote sensing (RS)” with “Remote sensing (hereafter named RS)”or similar.

Row 81  “scrutinize” seems to be not appropriated, pleas, change with a more suitable synonym

Row 343-345  Please, consider to rephrase.

Row 349  Please, replace “aprox.” with “approximatively”

Row 344  Please, explain which are the times and the reasons for the increase in the C. sativa spread.

Row 373-375  In the methods, it seems that the problem you talk about in these rows was solved by using another date for the Landsat 8 scene, but why here do it seem different?

Row 386-391  Please, consider to better argument you results also by adding literature.

Row 411  it is not clear how climatic changes can improve the spread of C. sativa.

Best regards 

 

Comments for author File: Comments.pdf

Author Response

Answers to Reviewer 1

Dear Reviewer #1

Thank you very much for your helpful review of our manuscript.  We are very pleased that you are positive about it! We hope we have responded appropriately to your comments; please see a point-by-point reply to each, below.

King regards,

The authors

________________________________________________

I really appreciated the manuscript, which use the new approach of RS and field data to assess the current distribution and the potential spread of the non-native plant Castanea sativa in island of La Palma.

 

I only suggest to better explain the approach of RS data analysis to people less confident with the use of RS. I'm not so familiar to give more precise suggestions.

Answer of the authors:

We extended the introduction with basic information about remote sensing application in invasion research and management. Also, according to other reviewer comments, we considerably revised and restructured the methods section. We think that our RS approach is now clearer. The changes made to the Introduction and Methods can be seen in the revised documents.

 

Please, check the authors names, there are some errors in the authors section and in the author contribution section.

Answer of the authors:

We apologize for that flaw. We are not sure what happened with the author names. Now the correct author names are listed in the manuscript and we hope that this is not in conflict with the data base that was created in parallel. In the author contribution section, we have now included all contributions of all authors.

 

Row 4-5  Please check the authors names.

Answer of the authors:

See our response to comment above.

 

Row 38-39 Please, delete the symbol of “~” from 1.2% and 4% and 13%

Answer of the authors:

Done.

 

Row 58  Please, delete “species intrusion” that is not appropriated to the Invasive Alien Species terminology, “threatened by ecological invasion” is enough.

Answer of the authors:

Done.

 

Row 72  Please, use the acronyms alone, by adding the list of acronyms at the beginning of the manuscript, or rather replace “Remote sensing (RS)” with “Remote sensing (hereafter named RS)”or similar.

Answer of the authors:

We decided on the second option.

 

Row 81  “scrutinize” seems to be not appropriated, pleas, change with a more suitable synonym Row 343-345  Please, consider to rephrase.

Answer of the authors:

This now reads “Following this development of RS approaches, we … “. (L )

 

Row 349  Please, replace “aprox.” with “approximatively”

Answer of the authors:

 We replaced it with “approximately”.

 

Row 344  Please, explain which are the times and the reasons for the increase in the C. sativa spread.

Answer of the authors:

Unfortunately, this request is not clear. It refers to an increase in the spread of chestnut, but we do not have a time-series that allows us to evaluate whether the rate of spread is increasing or decreasing, and we do not think such data currently exist. Probably the best route for getting such data is to develop a method for using the Landsat archive for that purpose, and we present that method in our manuscript.  A next step (for future research) is to apply the method to historical images.  Our empirical results in this paper simply inform about which parts of La Palma have appropriate environment for chestnut to spread. Beyond what we already state in the Introduction, therefore, we cannot at this stage state the rate of spread or explain the reasons for the invasion.

 

Row 373-375  In the methods, it seems that the problem you talk about in these rows was solved by using another date for the Landsat 8 scene, but why here do it seem different?

Answer of the authors:

The portion of the image which is contaminated with cloud (image from March 2017) was cropped out with cloud mask that is shipped as Quality Assessment (QA) band with Landsat 8 surface reflectance products and compensated the cropped portion with the image from (February 2017). Here we are pointing out that a small number of contaminated pixels remained on the edge of the image from March 2017 in the cropped area. This is now better explained in the manuscript.

 

Row 386-391  Please, consider to better argument you results also by adding literature.

Answer of the authors:

We added text and new references to this part of the discussion. Generally, the discussion was extended substantially, also following the recommendations of other reviewers. We added a series of new references to support this and other statements made in the discussion.

 

Row 411  it is not clear how climatic changes can improve the spread of C. sativa.

Answer of the authors:

This is a valid point as we are making a vague claim that climate change could promote the spread of C. sativa, which is just an assumption. And it is true that future climate change impact on C. sativa is not the scope of this study. Our study focuses on the current climatic conditions and the link between remote sensing detection of C.sativa combined with ground referencing and biogeographical modelling approaches. We identify currently suitable habitat conditions for the establishment of this tree species that is able to modify ecosystem structure, diversity and functioning. Future potential developments can be assessed with a following study on the basis of our findings but would make the paper too big and also too complex. In consequence, we deleted the section “in combination with climatic changes” in response to the point made by the Reviewer. In the near future, we plan to apply different climate models to this data set but this will be subject of a follow-up study.

 

Author Response File: Author Response.docx

Reviewer 2 Report

While I appreciate the authors’ efforts of mapping invasive species for ecological conservation in an island environment, I found a couple of major issues in the study. First, there is a lack of novelty from the remote sensing perspective. All the methods used are mature in the field, making the manuscript more like a lab report. I would suggest giving more emphasis to the new stuff if there is any. Second, I am not convinced by the super high accuracy (98.5% and 99.5%) based on such a routine classification approach. Invasive species detection in remote sensing remains a challenging task due to highly spectral similarities among species. The high accuracy in this study is probably due to the imbalance in sampling (presence vs. absence). Some details in the sample collection are missing though. It may also be related to the homogenous environment on the island. Then the question is how can the proposed framework be generalized to the other regions? Detailed comments follow.

• Section 2.2: You did not specify the dates of the images. It is also not clear how you dealt with the inconsistency in image resolution between Landsat and Sentinel.
• L131: How many field plots were deployed for C. santiva identification? What is the size of each plot? I am also concerned about the sample imbalance between the presence and the absence, which would lead to biased estimation and modeling results.
• L136: What are the spatial resolutions of the environmental variables? Do they have enough spatial variation over the island?
• L141-142: Shouldn't land cover be included as one of the explanatory variables?
• L140&L143: You used both r-squared=0.7 and r=0.7 as the threshold of high vs low correlation. But they are very different.
• L166: Feb to July is almost half a year. How did the seasonal difference affect your analysis?
• L195: Did you get two change images from three years? Be specific.
• Section 2.4: Did you apply RF to classify the band composition of D, d3, d2? You have three years of data. What change images did you apply RF to?
• L207: What is the accuracy of using Google Earth images as references to identify C. santiva?
• L213: Justify the use of 650 trees.
• L229: Why using 300 and 100m?
• Table 1: Your accuracy is surprisingly high. This goes back to the sampling imbalance concern.

Author Response

Dear Reviewer #2

Thank you very much for your valuable feedback! In response we made a number of changes, including re-running the SDM analysis with fewer pseudo-absence points. Below we respond specifically to each of your comments.

King regards,

The authors

________________________________________________

While I appreciate the authors’ efforts of mapping invasive species for ecological conservation in an island environment, I found a couple of major issues in the study. First, there is a lack of novelty from the remote sensing perspective. All the methods used are mature in the field, making the manuscript more like a lab report. I would suggest giving more emphasis to the new stuff if there is any.

Answer of the authors:

We are aware of the fact that we are using established RS approaches. However, the novelty in this paper is the combination of methods used: combining RS with ecological information (species phenology) and biogeographical modelling approaches. See also our next response, which explains why this is important. Our method enables us to identify the quality and also the limits of conventional RS approaches in the context of very steep terrain with inaccessible sites, where field work is simply impossible. In such cases RS is the only way to map vegetation patterns, but it is also hampered by the relief, shading, cloudiness etc. Thus we show how a combination of methods coming from different methodological background can improve the understanding of patterns such as in our case the spatial distribution and the options for invasion of C. sativa on the island of La Palma.

 

Second, I am not convinced by the super high accuracy (98.5% and 99.5%) based on such a routine classification approach. Invasive species detection in remote sensing remains a challenging task due to highly spectral similarities among species.

We cut and paste here a detailed comment from below because it is highly relevant to our response:

L166: Feb to July is almost half a year. How did the seasonal difference affect your analysis?

Answer of the authors to the two points:

We agree that invasive species detection with RS is usually challenging because of the spectral similarities. Indeed, we suggest that the fact we are able to overcome this problem with an elegantly simple solution is a key part of the novelty of the paper. As explained in the Abstract, Methods and other parts of the paper, chestnut is deciduous in La Palma (as elsewhere), while the forests it is invading are entirely evergreen. By selecting images in two different seasons – when chestnut is in leaf and when it is not in leaf, while the surrounding forest is in leaf throughout – we are able to clearly and very accurately identify where the species is.  This contrasts with other RS studies on invasive species, where this phenological distinction does not exist, and it explains the very high accuracy, in both relative and absolute terms.  It is also the reason why there is a difference of nearly half a year between the images: it is to sample different seasons to ensure that the chestnut is in leaf in one image and not in the other.  That is the point, and the effect on the analysis was to make it possible!  We think there is value – and novelty – in our demonstration of how very well this simple method works, and suggest that it is also applicable for other situations in which invasive species differ phenologically from the natives. Given that this was explained in both the abstract and other key parts of the paper, and the other three referees seem to have understood it, we are not sure how to make these points clearer, but we have gone through the manuscript to find ways of doing so (changes are tracked).  Please also see our next response.

 

The high accuracy in this study is probably due to the imbalance in sampling (presence vs. absence). Some details in the sample collection are missing though. It may also be related to the homogenous environment on the island.

Answer of the authors:

As explained in the previous response, we think that the high accuracy is primarily explained instead by the deciduous nature of the species invading evergreen vegetation. Further, the environment of the island is far from being homogeneous. As illustrated in the methods section, the island is often seen as a “mini-continent” with highly diverse climatic conditions ranging from almost desert like conditions to continuously humid laurel forest and from succulent scrub to the alpine zone. Even so, we take this point and have now redone the SDM analysis using a more balanced set of presence vs pseudoabsence data. This has resulted in slightly lower accuracy values, though they remain high for the reasons already stated.

 

Then the question is how can the proposed framework be generalized to the other regions?

Answer of the authors:

This study can be generalized to the other islands of the archipelago, where comparable climatic conditions occur and those are responded naturally by the characteristic ecosystem of the laurel forest. These are the islands El Hierro, La Gomera, Tenerife, and Gran Canaria. There the findings of the study can be used as an early warning to generate awareness of possible processes and proactive measures can be undertaken to avoid invasion into these valuable remnant forests that are unique. Additional transfer can be done to the islands of Madeira and the Azores, where climatic conditions are very likely even more appropriate to C. sativa. It is very likely to be true of other oceanic archipelagos, too, and we suggest that phenological differences between natives and invaders may in fact be found in other places, too.  All these things need to be investigated with further studies. In the case of the Atlantic archipelagos, we also note that in the Azores the laurel forest is almost completely replaced by conifer plantations and other invasive species (e.g. Pittosporum undulatum, Hedychium gardnerianum). This makes the preservation of the Canary Island laurel forest an even more important priority in the international context! We added these arguments to the discussion.

 

   Detailed comments follow.

Section 2.2: You did not specify the dates of the images. It is also not clear how you dealt with the inconsistency in image resolution between Landsat and Sentinel.

   Answer of the authors:

The details about images are in the appendix and we have referenced that in the main text, but now we have now added the information in the main text as well.

Image resolution: Landsat 8 and Sentinel-2 data were analysed separately. For Landsat, we simply compared two images (different seasons) in their original resolution, and the same for Sentinel-2. So in the analyses there was no inconsistency to deal with! The one part of the manuscript for which this is an issue is in Figure 5, where we downscaled the Landsat data to 10m resolution in order to make the comparison between the results in terms of numbers of pixels.  We hope this is now clearer in the revised manuscript.

 

 L131: How many field plots were deployed for C. sativa identification? What is the size of each plot? I am also concerned about the sample imbalance between the presence and the absence, which would lead to biased estimation and modeling results.

   Answer of the authors:

We didn’t rely on the field collected data for C. sativa detection – that is the reason why we applied the change detection and supervised classification technique. We developed the change detection technique utilizing the phenological cycle of the species to identify it, and trained pixels with C. sativa and without C. sativa, and ran supervised classification to discern C. sativa. Our field-collected data are just for reference to check that the detected changes match C. sativa in the field.  We hope these points are now clearer in the manuscript.

 

 L136: What are the spatial resolutions of the environmental variables? Do they have enough spatial variation over the island?

   Answer of the authors:

The spatial resolution of the environmental variables is 100 m * 100 m.

La Palma has very steep relief and rises from the sea to around 2300 m in elevation. This translates into steep gradients of environmental parameters. It also has a distinct rain shadow and a strong gradient of precipitation determined in large part by interception at mid-altitudes of moisture from the trade winds, with low and high elevations much drier.  Indeed, the variation in environmental parameters makes the island very attractive for studying responses of species along these gradients, which is one reason why there are many ecological studies on the island.

 

 L141-142: Shouldn't land cover be included as one of the explanatory variables?

    Answer of the authors:

This is difficult because the CORINE land cover classes do not really reflect either the land use or the ecosystem types of the Canary Islands, because these classes were made for continental Europe. However, we have prepared a map (shape file) for the main ecosystem considered in this paper, which is the laurel forest. This map is based on the in-situ mapping of highly differentiated plant communities under the authority of the regional government (Cabildo Insular). As these mapping units are too differentiated, we aggregated those of the laurel forest to show the extent of this ecosystem in our maps. Nevertheless, such maps cannot be used as explanatory variables because they are just based on expert knowledge but not on measured and precisely localised data. The shape file of the laurel forest has been included in the SDM maps that are projecting the potential distribution of C. sativa in order to visualize the overlap.

 

L140&L143: You used both r-squared=0.7 and r=0.7 as the threshold of high vs low correlation. But they are very different.

   Answer of the authors:

Thanks for spotting this. The instances of ‘r2’ were typos, and should have simply been ‘r’. We have now corrected this.

 

 L166: Feb to July is almost half a year. How did the seasonal difference affect your analysis?

    Answer of the authors:

This comment is pasted above and answered there.

 

 L195: Did you get two change images from three years? Be specific.

   Answer of the authors:

We generated change images by ourselves. No, they are not from three years. They are from two different sensors i.e. Landsat 8 and Sentinel-2. We used images from each sensor from different seasons. We took images from different seasons because the comparison of the images from two seasons makes possible to detect the species.

 

 Section 2.4: Did you apply RF to classify the band composition of D, d3, d2? You have three years of data. What change images did you apply RF to?

   Answer of the authors:

Yes, we applied RF to classify the band composition of D, d3, and d2. Those bands were obtained from the change detection images generated by ourselves. We made comparison between different seasons. Landsat 8 from March 2017 and July 2017 for Landsat 8 and July 2018 and February 2019 for Sentinel 2. We used images from different seasons to generate change map.

 

 L207: What is the accuracy of using Google Earth images as references to identify C. santiva?

   Answer of the authors:

Google Earth was used to add locations in clearly attributed chestnut forests that can be clearly differentiated when looking at different months of the year because of the deciduous nature of the species.

 

 L213: Justify the use of 650 trees.

    Answer of the authors:

The OOB error in the random forest classification was near-constant after that (i.e. when using more than 650 trees), so we used 650.

 

L229: Why using 300 and 100m?

    Answer of the authors:

 We used 100 m-thinning for field data and 300 m-thinning for RS data because distances between irregularly distributed field records were often much smaller than distances between gridded RS data. With respect to the spatial resolution of RS data, 300 m was considered appropriate, but less than 300 m distance was considered appropriate for field records.

 

Table 1: Your accuracy is surprisingly high. This goes back to the sampling imbalance concern.

Answer of the authors:

As stated above, we have addressed the sampling imbalance concern by rerunning the analysis with a more balanced sample.  This only slightly reduces the accuracy, which is mainly due to the clear distinction of the deciduous tree species versus the entirely evergreen tree species without seasonal fluctuations in their matrix. Please see above for a fuller response on this issue.

 

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript “Detecting the replacement of laurel forest by a novel ecosystem in the steep terrain of an oceanic island” is assessing the potential of non-native Castanea sativa Mill. and its spreading into native and
endemic ecosystems of La Palma. It deals with one of  the very important problems of biodiversity, such as replacement of native by non-native plant communities. Results are valuable and maybe of international interest. The authors demonstrated a good knowledge of the problematic and of the related literature. The paper is well structured.

Nonetheless, still some minor reviews should be performed.

Please check lines 28, 61, and 62. In accordance with scientific standards and International Code of Botanical Nomenclature (ICBN), scientific plant names, irrespective of rank, should be given in italics. The author’s name should be written in normal print at least once, when mentioned for the first time in the text or in a table, and should be omitted subsequently. They should be abbreviated in conformity with the Authors of Plant Names, easily accessible on https://www.ipni.org/ or http://www.theplantlist.org/. After the first mention, the generic name should be abbreviated to its initial. 

Line 81: Remove (https://doi.org/10.1111/avsc.12403).

Use the term "In situ" correctly.When used as an adverb, there is no hyphen; when used as an adjective, there is a hyphen.

In my opinion, the introduction is nice, but a little short. I would advise the authors to expand it in order to make the manuscript excellent.

Author Response

Dear Reviewer #3

Thank you very much for your review! We appreciate how positive you are about our paper, and also your suggestions for improvement, which are valuable.

King regards,

The authors

________________________________________________

The manuscript “Detecting the replacement of laurel forest by a novel ecosystem in the steep terrain of an oceanic island” is assessing the potential of non-native Castanea sativa Mill. and its spreading into native and endemic ecosystems of La Palma. It deals with one of  the very important problems of biodiversity, such as replacement of native by non-native plant communities. Results are valuable and maybe of international interest. The authors demonstrated a good knowledge of the problematic and of the related literature. The paper is well structured.

Nonetheless, still some minor reviews should be performed.

 

Please check lines 28, 61, and 62. In accordance with scientific standards and International Code of Botanical Nomenclature (ICBN), scientific plant names, irrespective of rank, should be given in italics.

Answer of the authors:

This was corrected. All taxa were checked for accepted taxonomy in World Flora Online which is linked to IPNI and ICBN.

 

The author’s name should be written in normal print at least once, when mentioned for the first time in the text or in a table, and should be omitted subsequently. They should be abbreviated in conformity with the Authors of Plant Names, easily accessible on https://www.ipni.org/ or http://www.theplantlist.org/. After the first mention, the generic name should be abbreviated to its initial.

Answer of the authors:

Done.  As “The Plant List” is no longer updated since 2013, we refer to “World Flora Online”
(http://www.worldfloraonline.org/) and IPNI, respectively. We checked plant taxa with this data base and applied correct nomenclature of accepted taxa throughout the manuscript. We added attribution (e.g. lines 71, 112, 118).

 

Line 81: Remove (https://doi.org/10.1111/avsc.12403).

Answer of the authors:

 Done.

 

Use the term "In situ" correctly. When used as an adverb, there is no hyphen; when used as an adjective, there is a hyphen.

Answer of the authors:

 Done. We corrected in one case to “in situ”.

 

In my opinion, the introduction is nice, but a little short. I would advise the authors to expand it in order to make the manuscript excellent.

Answer of the authors:

We integrated additional sentences into the introduction, about invasive species and about the progress in  invasion research due to remote sensing applications, also critically pointing at limits and restrictions in several of these studies. In the discussion we pointed at the future application of other RS approaches such as the use hyperspectral sensors that are made available also via satellites. However, we focus in our study on the currently available RS products and on a case where these are working well.  Please see the very substantial changes to the Introduction and Discussion that are tracked in the revised manuscript.

 

Author Response File: Author Response.docx

Reviewer 4 Report

The study describes an attempt for mapping the habitat of a potential invasive species such as the chestnut (Castanea sativa) in La Palma (Canary Islands, Spain), using a model-assisted framework and based on in-field and remote sensing data from two satellites, Landsat and Sentinel. Firstly, the authors extracted seasonal images considering the distinctive phenological behaviour of the target species to detect spectral changes between native and alien species by image regressions. They then applied a supervised classification to map the present-day occurrence of the target species. Authors also calibrated a habitat suitability model through three algorithms based on both in-field and remote sensed occurrences and using a set of environmental predictors, to detect potential invasion areas. Finally, they compared the model performance and identified if satellite-derived data best explain the potential distribution of the alien species through the study area.

 

Overall, the paper touches a topical subject in the context of detecting the expansion of a potential alien species through a native forest and using a multi-technique approach. All sections are generally concise and stringent. Figures and tables are usually accurate and useful, but the quality of figures, in terms of color ramps and self-explanation, must be improved. Also, they should be cited by keeping the order of appearance in the text. Thus, and unfortunately, there are enough reasons not to accept the article in its current form, since it requires a major review of many of the points discussed and further improvements. For instance, the title is too generic, it does not represent well the objectives addressed in this study and it could lead to a misinterpretation. I recommend focussing on how the potential distribution of the target species will be, considering the different data sources used, and not so much in the replacement since this has not been tested empirically. I also recommend a reorganization of the sections and subsections, especially in Methods and Results. In addition, despite there were cited examples on the use of remote sensing data to mapping (alien)species, the references list is not enough supported by case studies/broad literature. The structure of author´s names and affiliations should be checked, in the main text and in supplementary material. Following are my comments and suggestions.

 

General comments

1. The introduction is well balanced in length and correctly depicts the state of the art and background, but needs a broader literature on further research or approaches (e.g. by using other remote sensing sensors and techniques/approaches) to mapping habitat suitability. In addition, the references style should be checked since it does not match the journal style.
2. Methods and Results sections are in general well described, but please see specific comments. In general, the workflow was well described, but an improvement is needed in terms of appearance and to be more explicit in technical matters to guarantee the replicability of this study.
3. Discussion section is also poorly and vague stated. Despite its promising results, the authors have not faced its methodology against other approaches, and neither considering or testing other methods or algorithms/classifiers of habitat modelling that consider presence data. Key literature on further research or approaches needs to be reviewed (e.g. more general spatial and temporal predictive models’ applications, more emphasis on environmental descriptors that also inform on the structural/functional component of the ecosystems, and so on). This could be solved by providing appropriate references to reinforce their statements. Thus, a more general review on the application of different remote sensing data and approaches to predict species invasion/habitat in space and time, taxa, etc [see 4-8 references below as examples, among other], should be addressed.

 

Specific comments:

- Line 81. Check reference style and remove doi.

- Line 116. “… zoom …” instead of “… magnifications …”?

- Line 136. If predictors were processed to match spatial resolutions of both source data of species occurrences, in-field and remote sensing, please stated it in detail somewhere.

- Line 165-167. It is not clear if the authors used an image collection for both two periods, or just single-monthly images.

- Line 228. Please keep the order of appearance.

- Line 233. Why thousands? Please check this reference [1].

- Lines 227-236. This paragraph should be moved to the data section, specifically, the species data section. I recommend generating a few more subsections starting with the description of the study area, species data, predictors, change detection, ENMs, and so on, and keep this scheme for Methods and Results.

- Line 237. Not clear if this is the right metric used by biomod2:

• The total sum of squares, or SST, is a measure of the variation of each response value around the mean of the response. For each observation, this is the difference between the response value and the overall mean response.
• The true skill statistics, or TSS, is defined based on the components of the standard confusion matrix representing matches and mismatches between observations and predictions.

Please check and confirm that you are using the right metrics, or just provide the correct definition to demonstrate that they are being correctly using by the authors. Otherwise this could raise doubts.

- Line 239. “… because TSS and the ROC scores were found better …” What are the authors based on to say this? Please explain or provide a reference.

- Line 243. It is better to move this figure 8 to Results if you are going to perform a habitat suitability assessment. Again, it is recommended to cite figures in order of appearance.

- Line 244. Idem.

- Line 249. This statement is too optimistic considering the results and watching the figure. Please change color ramp because it is hard to find these agreements. Rewrite and check comments on figures below.

- Lines 250-253. One way to check that would be to face the surface occupied by predicted pixels of these ambiguities’ areas against land cover pixels and confirm it.

- Lines 278-279. This is already stated above. Maybe the authors could move or merge some paragraphs.

- Lines 279-280. Idem. Figure 5 has not yet been cited.

- Line 295. “… (blue), Landsat 8 and Sentinel-2 both (black), and …” Shared or added? Please be more explicit.

- Lines 295-296. I have doubts about this test was performed correctly. The regular way is to aggregate pixels at fine resolution to broader resolution, because if a pixel at 30 m is downscaled to 10 m, the number of pixels for Landsat will irretrievably increase, and that would have an effect on increasing the occupied surface. So please provide additional explanation or rewrite correctly.

- Line 297. This table is not cited until the line 321, so please keep the order of appearance.

- Lines 297-298. To ensure the replicability of this study, these classification metrics should be described, at least succinctly, in Methods section. For instance, describe what is “OOB” for the broadly audience.

- Line 301. “… in the images from the Landsat 8 and Sentinel-2, …” Overlapping?

- Line 302. “… zoom …” instead of “… magnifications …”?

- Lines 302-303. Not sure if these types of sentences within the figures are needed, it is obvious.

- Lines 305-308. Perhaps the authors can explain this info just once within the Methods section and not repeat them in every legend.

- Line 311. Which models? Here is recommended to remind the acronym of each model, or just to add “… ENM models …”.

- Lines 311-312. This sentence is hard to read. Please rewrite or used some punctuation mark.

- Line 315. Perhaps Fig. 7?

- Lines 315-318. That matches with the observed in the field?

- Line 321. Perhaps Table 2?

- Line 321. The cross-validation outputs use to be shown before the map’s analysis.

- Line 324. Remove the comma after “… a) …”

- Line 338. The potential of what? To invade? I think the authors are only assessing the current and potential distribution of the target species into native ecosystems, not the invasiveness of species. Please rewrite.

- Line 340. It is recommended to provide somehow the distribution of this type of forest to see if there is an overlap, as it was not empirically tested.

- Lines 357-358. No data or results are provided to support this statement.

- Lines 362. “… two consecutive dates …”. Please provide any reference. See below.

- Lines 365-367. That is obvious and does not contribute anything new. I recommend reviewing similar studies where the effect of pixel size between different sensors and different dates is discussed. Some recommendations at the bottom of this review.

- Lines 368-370. A smaller predicted spatial coverage does not have to depend only on a smaller pixel size, but on the information that the sensor captures in each pixel. Perhaps Sentinel is more conservative when it comes to predicting pixels, but it would be necessary to test whether the spectral/environmental information captured by this sensor is similar to that of Landsat, which is more generalist in prediction.

Lines 373-374. Please check reference [3].

- Line 375. It does not have to be if the atmospheric correction has been done well.

- Line 377-379. The authors could confirm, and therefore reinforce, these statements based on the response curves or variable importance derived from the ENMs.

- Line 380. Diverse in what way?

- Lines 385-386. Hard to understand by broader audience. Try to provide references. E.g., [2].

- Lines 386-388. Any reference to support this statement?

- Lines 400-401. Please consider this statement and support with a deeper search on related bibliography.

- Line 407. “…Only in valley bottoms two rare deciduous trees can be found naturally. …” This sentence is hard to locate in the narrative. Please rewrite or be more explicit.

 

Figure 1. To improve the self-explanation of figures, please provide different dates for each image (boxes).

Figure 2. Please change the color ramp for changed pixels because there is not a good contrast with the background.

Figure A (SM). Please change color ramp.

Table 2. How and why these thresholds were choosing? In addition, could the area be the other way around?

 

Additional references suggested in this review.

[1] Manel, S., Williams, H.C. and Ormerod, S. (2001), Evaluating presence–absence models in ecology: the need to account for prevalence. Journal of Applied Ecology, 38: 921-931. doi:10.1046/j.1365-2664.2001.00647.x

[2] Mateo RG, Mokany K, Guisan A. Biodiversity Models: What If Unsaturation Is the Rule?. Trends Ecol Evol. 2017;32(8):556-566. doi:10.1016/j.tree.2017.05.003

[3] Arenas-Castro, S.; Fernández-Haeger, J.; Jordano-Barbudo, D. Evaluation and Comparison of QuickBird and ADS40-SH52 Multispectral Imagery for Mapping Iberian Wild Pear Trees (Pyrus bourgaeana, Decne) in a Mediterranean Mixed Forest. Forests 2014, 5, 1304-1330.

[4] Andrew, M.E. and Ustin, S.L. (2009), Habitat suitability modelling of an invasive plant with advanced remote sensing data. Diversity and Distributions, 15: 627-640. doi:10.1111/j.1472-4642.2009.00568.x

[5] Bolch E.A. et al. (2020) Remote Detection of Invasive Alien Species. In: Cavender-Bares J., Gamon J., Townsend P. (eds) Remote Sensing of Plant Biodiversity. Springer, Cham. https://doi.org/10.1007/978-3-030-33157-3_12

[6] Huang CY, Asner GP. Applications of remote sensing to alien invasive plant studies. Sensors (Basel). 2009;9(6):4869-89. doi: 10.3390/s90604869. Epub 2009 Jun 19. PMID: 22408558; PMCID: PMC3291943.

[7] Rocchini D, Andreo V, Förster M, et al. Potential of remote sensing to predict species invasions: A modelling perspective. Progress in Physical Geography: Earth and Environment. 2015;39(3):283-309. doi:10.1177/0309133315574659

[8] Vaz AS, Alcaraz-Segura D, Vicente JR and Honrado JP (2019) The Many Roles of Remote Sensing in Invasion Science. Front. Ecol. Evol. 7:370. doi: 10.3389/fevo.2019.00370

Author Response

Dear Reviewer #4

Thank you very much for your valuable feedback! Based on your comments, we could improve the manuscript a lot. We responded to each of your comments below.

King regards,

The authors

________________________________________________

The study describes an attempt for mapping the habitat of a potential invasive species such as the chestnut (Castanea sativa) in La Palma (Canary Islands, Spain), using a model-assisted framework and based on in-field and remote sensing data from two satellites, Landsat and Sentinel. Firstly, the authors extracted seasonal images considering the distinctive phenological behaviour of the target species to detect spectral changes between native and alien species by image regressions. They then applied a supervised classification to map the present-day occurrence of the target species. Authors also calibrated a habitat suitability model through three algorithms based on both in-field and remote sensed occurrences and using a set of environmental predictors, to detect potential invasion areas. Finally, they compared the model performance and identified if satellite-derived data best explain the potential distribution of the alien species through the study area.

 

Overall, the paper touches a topical subject in the context of detecting the expansion of a potential alien species through a native forest and using a multi-technique approach. All sections are generally concise and stringent. Figures and tables are usually accurate and useful, but the quality of figures, in terms of color ramps and self-explanation, must be improved.

Answer of the authors:

We revised the maps and modfied the color schemes.

Maps displaying the results of environmental niche modelling were updated with new modelling results and their appearance improved. Additionally, the ecological information of the maps was increased by including information on the distribution of the laurel forest on La Palma (Fig. 8).

 

Also, they should be cited by keeping the order of appearance in the text.

Answer of the authors:

The order and appearance of all figures and tables was checked and changed when necessary so that they now appear in the numerically and alphabetically correct order.

 

Thus, and unfortunately, there are enough reasons not to accept the article in its current form, since it requires a major review of many of the points discussed and further improvements. For instance, the title is too generic, it does not represent well the objectives addressed in this study and it could lead to a misinterpretation. I recommend focussing on how the potential distribution of the target species will be, considering the different data sources used, and not so much in the replacement since this has not been tested empirically.

Answer of the authors:

We modified the title to avoid misinterpretation. “Detecting the replacement of laurel forest by a novel ecosystem …" was replaced by “Assessing the potential replacement of laurel forest by a novel ecosystem …".

However, we kept laurel forest in the title because the preservation and threat of this unique ecosystem is of interest, whereas chestnut itself is of less relevance, if it would not have the potential to replace the laurel forest. The current distribution with some vast forests dominated by chestnut in the direct vicinity of laurel forest has evoked the necessity for this study which was initiated by practitioners in the environmental administration of the island.

 

I also recommend a reorganization of the sections and subsections, especially in Methods and Results.

Answer of the authors:

The Methods were reordered and now consist clearly of one part about ‘Change detection’ and a second part about ‘Environmental niche modelling’. By revising the Results, including the order of tables and figures, we think they now appear in a clearer and more logical order.

 

In addition, despite there were cited examples on the use of remote sensing data to mapping (alien)species, the references list is not enough supported by case studies/broad literature.

Answer of the authors:

Additional references were integrated, particularly during the revision and extension of the Introduction and Discussion chapters. For instance, all eight papers suggested by Reviewer #4 (among many more) are now embedded into the main text. In addition, we added more explicit explanations and discussions on the use of RS in invasion research including the application of different kind of sensors as well airborne as satellites with the respective references in order to position our work better in the frame of the general research landscape. Also we added papers that are specifically detecting tree species in different ecosystems.

 

The structure of author´s names and affiliations should be checked, in the main text and in supplementary material.

Answer of the authors:

Unfortunately, we cannot explain how this happened. The names were correct in the file we submitted. We cannot exclude some kind of system fault during submission. We apologize for that and hope the files retain the correct author names this time.

 

Following are my comments and suggestions.

 General comments

    The introduction is well balanced in length and correctly depicts the state of the art and background, but needs a broader literature on further research or approaches (e.g. by using other remote sensing sensors and techniques/approaches) to mapping habitat suitability.

Answer of the authors:

We considerably extended the Introduction and Discussion, adding literature on various remote sensing applications in detecting and mapping invasive species. There, we also show now what has been done in previous studies, pointing out that these were often done with non-public and cost-intensive very high resolution and airborne sensors. We also show examples of uses of other satellite sensors that may have limited access to data, or lifetime of the satellite. This now illustrates why a study using for instance Sentinel multi-temporal data that are free to use is of relevance for applied studies on invasive species.

 

In addition, the references style should be checked since it does not match the journal style.

    Answer of the authors:

Done.

 

Methods and Results sections are in general well described, but please see specific comments. In general, the workflow was well described, but an improvement is needed in terms of appearance and to be more explicit in technical matters to guarantee the replicability of this study.

   Answer of the authors:

The workflow of the methods and results sections has been adjusted to make the single steps of this paper comprehensible. Furthermore, the appearance of figures and tables has been adjusted so that they now follow a more logical order. These two adjustments, together with the inclusion of further technical details, have improved the replicability of the study.

 

 Discussion section is also poorly and vague stated. Despite its promising results, the authors have not faced its methodology against other approaches, and neither considering or testing other methods or algorithms/classifiers of habitat modelling that consider presence data.

Answer of the authors:

We greatly extended the discussion, contrasting our approach with previous studies using different sensors, classification algorithms, habitat suitability models and model evaluation techniques. We thus highlight the challenge of selecting RS data and methodological approaches to yield highest detection and prediction success. Please see the tracked changes in the revised Discussion.

 

Key literature on further research or approaches needs to be reviewed (e.g. more general spatial and temporal predictive models’ applications, more emphasis on environmental descriptors that also inform on the structural/functional component of the ecosystems, and so on). This could be solved by providing appropriate references to reinforce their statements. Thus, a more general review on the application of different remote sensing data and approaches to predict species invasion/habitat in space and time, taxa, etc [see 4-8 references below as examples, among other], should be addressed.

 Answer of the authors:

We appreciate the literature suggestions a lot - thank you very much! As mentioned above, we added several paragraphs that review and discuss the applications of other RS data, image classification techniques and modelling approaches. In the process, we incorporated all the references suggested, as well as additional ones that are related to the recommended extension of the introduction and the discussion.

 

Specific comments:

 - Line 81. Check reference style and remove doi.

 Answer of the authors:

Done.

 

- Line 116. “… zoom …” instead of “… magnifications …”?

 Answer of the authors:

It has been changed to “zooms”.

 

- Line 136. If predictors were processed to match spatial resolutions of both source data of species occurrences, in-field and remote sensing, please stated it in detail somewhere.

 Answer of the authors:

The environmental variables are in 100 m spatial resolution except elevation, slope and aspect. We used elevation data of 2 m spatial resolution and calculated slope and aspect in QGIS. We then resampled the data (elevation, aspect, and slope) to 100m spatial resolution.

We added the following sentence: “A spatial resolution of 100 m was given for all environmental variables except elevation, slope and aspect; we aggregated the resolution of elevation, slope and aspect to 100 m.”

 

- Line 165-167. It is not clear if the authors used an image collection for both two periods, or just single-monthly images.

 Answer of the authors:

The Landsat images were taken from two dates (one form March 07, 2017 and another from July, 29 2017) for analysis and one from February 03, 2017 to compensate cloud. Similarly, two images from Sentinel –2 (from July 08, 2018 and February 13, 2019). We have also mentioned the images’ IDs in the appendix, and we cited the appendix in the main text. Now we have also mentioned the date in main text.

 

- Line 228. Please keep the order of appearance.

 Answer of the authors:

Figures and tables are now mentioned in order of appearance.

 

- Line 233. Why thousands? Please check this reference [1].

 Answer of the authors:

We accounted for prevalence (Manel et al. 2001 & Liu et al. 2005), changed the number of pseudo-absences accordingly and redid the model. We now use the same number of pseudo absences as presences (as advocated by Liu et al. 2005). Additionally, we increased the quality of the modelling outcome by increasing the sets of pseudo absences from 3 to 10 and by increasing the modelling repetitions from 3 to 4.

 

- Lines 227-236. This paragraph should be moved to the data section, specifically, the species data section. I recommend generating a few more subsections starting with the description of the study area, species data, predictors, change detection, ENMs, and so on, and keep this scheme for Methods and Results.

Answer of the authors:

We restructured the Methods section considerably. Please see tracked changes in the revised manuscript. It is now clearly divided in a remote sensing and an environmental niche modelling part. We therefore did not remove this paragraph somewhere else to make the reader understand that these species data were only used for environmental niche modelling and not for analysis related to remote sensing. However, we agree that the methods needed restructuring and we hope the new version fulfils your requirements.

 

- Line 237. Not clear if this is the right metric used by biomod2:

    The total sum of squares, or SST, is a measure of the variation of each response value around the mean of the response. For each observation, this is the difference between the response value and the overall mean response.

    The true skill statistics, or TSS, is defined based on the components of the standard confusion matrix representing matches and mismatches between observations and predictions.

Please check and confirm that you are using the right metrics, or just provide the correct definition to demonstrate that they are being correctly using by the authors. Otherwise, this could raise doubts.

 Answer of the authors:

Thank you for picking this up. We checked and adjusted the term to “true skill statistics”. Furthermore, we checked for correct use of the model evaluation metric and confirm that we used true skill statistics to determine our model performance.

 

- Line 239. “… because TSS and the ROC scores were found better …” What are the authors based on to say this? Please explain or provide a reference.

 Answer of the authors:

We have modifed our modelling. Now we used all the six outputs. They are weighted sum of probabilities, committee averaging across predictions, and mean probabilities across predictions.

 

- Line 243. It is better to move this figure 8 to Results if you are going to perform a habitat suitability assessment. Again, it is recommended to cite figures in order of appearance.

 Answer of the authors:

Figure 8 has been moved to the results and the figures are now cited in order of appearance.

 

- Line 244. Idem.

 Answer of the authors:

The figures are now cited in order of appearance.

 

- Line 249. This statement is too optimistic considering the results and watching the figure. Please change color ramp because it is hard to find these agreements. Rewrite and check comments on figures below.

 Answer of the authors:

We adjusted the colour ramp and now write “large spatial agreement” instead of “excellent spatial agreement”.

 

- Lines 250-253. One way to check that would be to face the surface occupied by predicted pixels of these ambiguities’ areas against land cover pixels and confirm it.

 Answer of the authors:

The available land cover classifications do not apply well to the conditions on La Palma and the Canaries as CORINE is designed for the European continent.

 

- Lines 278-279. This is already stated above. Maybe the authors could move or merge some paragraphs.

 Answer of the authors:

We agree and removed this sentence.

 

- Lines 279-280. Idem. Figure 5 has not yet been cited.

 Answer of the authors:

The issue has been resolved and Figure 5 is now cited in the text.

 

- Line 295. “… (blue), Landsat 8 and Sentinel-2 both (black), and …” Shared or added? Please be more explicit.

 Answer of the authors:

It is shared by both the sensors. We have clarified this in the main text.

 

- Lines 295-296. I have doubts about this test was performed correctly. The regular way is to aggregate pixels at fine resolution to broader resolution, because if a pixel at 30 m is downscaled to 10 m, the number of pixels for Landsat will irretrievably increase, and that would have an effect on increasing the occupied surface. So please provide additional explanation or rewrite correctly.

 Answer of the authors:

Yes, either downscaling or upscaling would change the number of pixels. We did this simply for the purposes of comparison. Our aim was to compare the pixels of two sensors against elevation. For example, in a 120 m* 120 m block there are 4*4 pixels of Landsat 8 and 12*12 pixels of Sentinel-2. So, regardless of downscaling Landsat 8 or upscaling Sentinel-2 the area remains equal. Our aim is to see what are the areas that each of the sensors and the area they share occupy at different altitude. In other words, we want to check the performance of each sensor. Therefore, changes in the number of pixels doesn’t affect our purpose.

However, the issue of change detection and pixel size is now addressed more in detail.

As these two sensors are widely used we are convinced that the comparison makes sense to readers. However, there can be a more general question behind this remark of the reviewer that is valid: Downscaling Landsat to 10m pixels could tend to over-estimate area relative to Sentinel, while upscaling Sentinel to 30m pixels would not have such a bias, and would therefore be a more appropriate way to compare the areas. If this is meant, then we cannot address this point. Such a general issue of up- and downscaling exists in different kinds of spatial data sets including Red Listing (see Moat, Bachman, Field & Boyd 2018 Conservation Biology 32, 1278–1289) creating bias in whichever way it is done, but stronger when downscaling. Nevertheless, we think that this is a very general problem that cannot be discussed on a theoretical basis in the frame of this paper but would be worth a review or methodological study.

 

- Line 297. This table is not cited until the line 321, so please keep the order of appearance.

 Answer of the authors:

We changed the order of appearance accordingly.

 

- Lines 297-298. To ensure the replicability of this study, these classification metrics should be described, at least succinctly, in Methods section. For instance, describe what is “OOB” for the broadly audience.

 Answer of the authors:

We have updated the information in the Methods.

 

- Line 301. “… in the images from the Landsat 8 and Sentinel-2, …” Overlapping?

 Answer of the authors:

We added “overlapping”.

 

- Line 302. “… zoom …” instead of “… magnifications …”?

Answer of the authors:

 As requested below, these kinds of figure descriptions were removed.

 

- Lines 302-303. Not sure if these types of sentences within the figures are needed, it is obvious.

Answer of the authors:

 See comment above. Fair point; we removed this information.

 

- Lines 305-308. Perhaps the authors can explain this info just once within the Methods section and not repeat them in every legend.

Answer of the authors:

 We would like to preserve this info on forests and natural systems because we think it is helpful for the reader to understand the figures without referring to the main text.

 

- Line 311. Which models? Here is recommended to remind the acronym of each model, or just to add “… ENM models …”.

Answer of the authors:

 We have updated the main text to “The ENMs showed...”

 

- Lines 311-312. This sentence is hard to read. Please rewrite or used some punctuation mark.

Answer of the authors:

 The sentences now read “All ENMs showed that habitats in the eastern and northern parts of the island - including the areas of present distributions - were more suitable for C. sativa (Fig. 7 and 8, Appendix F and G for single model results). The ENMs based on species occurrences from field observation and the ENMs based on species occurrences from RS data were found to have very good AUCs and TSS scores (Table 2 and Appendix H).”

 

- Line 315. Perhaps Fig. 7?

Answer of the authors:

 Yes! We have updated the main text: “(Fig. 7a and 8a)”

 

- Lines 315-318. That matches with the observed in the field?

Answer of the authors:

 We are not sure what this question aims at. Here we simply compare the predicted habitat suitability based on the field data and remote sensing data.

 

- Line 321. Perhaps Table 2?

Answer of the authors:

Yes, we have corrected this information.

 

- Line 321. The cross-validation outputs use to be shown before the map’s analysis.

Answer of the authors:

Yes.  We have now rearranged the sequence.

 

- Line 324. Remove the comma after “… a) …”

Answer of the authors:

 Done.

 

- Line 338. The potential of what? To invade? I think the authors are only assessing the current and potential distribution of the target species into native ecosystems, not the invasiveness of species. Please rewrite.

Answer of the authors:

You are right! We rephrased the text accordingly: “This study is assessing the potential current and potential distribution of non-native C. sativa and the spread into, invading native and endemic ecosystems of La Palma. “

 

- Line 340. It is recommended to provide somehow the distribution of this type of forest to see if there is an overlap, as it was not empirically tested.

Answer of the authors:

We added a shape file on the current distribution of laurel forest. Differentiated plant communities were mapped by official administration (Cabildo Insular) on the basis of the almost 400 plant communities that were published in this context in the book “Vegetation of the Canary Islands” (Del Arco Aguillar & Delgado 2018, Springer). However, as these units are much too differentiated, we aggregated all laurel forest communities to the laurel forest ecosystem for this island under consideration. This addition to the map (yellow colour) now shows the extent and overlap of laurel forest with the projected potential distribution of C. sativa.  

 

- Lines 357-358. No data or results are provided to support this statement.

Answer of the authors:

We revised Figure 8. We think this statement is now supported by this map including both the predicted niche of C. sativa and the current distribution of laurel forest. The maps now illustrate that C. sativa first of all could affect the marginal areas of the laurel forest and thus reduce its area to a small core that might not suffice to maintain the unique diversity of this ecosystem.

 

- Lines 362. “… two …”. Please provide any reference. See below.

Answer of the authors:

We have rewritten the sentence as follows: “This change detection technique is especially suitable for invasive plant species detection if the species exhibits clear phenological changes between two consecutive dates compared to native vegetation through time, as shown by Labonté al. (2020) detecting glossy buckthorn (Frangula alnus Mill.) spreading into forests of southern Quebec, Canada, by using a linear temporal unmixing model to a time series of the normalized difference vegetation index (NDVI) derived from Landsat 8 Operational Land Imager (OLI) images.”

 

- Lines 365-367. That is obvious and does not contribute anything new. I recommend reviewing similar studies where the effect of pixel size between different sensors and different dates is discussed. Some recommendations at the bottom of this review.

Answer of the authors:

We added several paragraphs discussing the effect of pixel size and acquisition dates on detection and prediction accuracy through similar studies. Please see tracked changes in the Discussion.

 

- Lines 368-370. A smaller predicted spatial coverage does not have to depend only on a smaller pixel size, but on the information that the sensor captures in each pixel. Perhaps Sentinel is more conservative when it comes to predicting pixels, but it would be necessary to test whether the spectral/environmental information captured by this sensor is similar to that of Landsat, which is more generalist in prediction.

Answer of the authors:

 According to https://www.usgs.gov/centers/eros/science/usgs-eros-archive-sentinel-2-comparison-sentinel-2-and-landsat?qt-science_center_objects=0#qt-science_center_objects Sentinel-2A data have spectral bands very similar to Landsat 8. We thus assume that the spectral information captured by Sentinel-2A is comparable to Landsat 8.

 

-Lines 373-374. Please check reference [3].

Answer of the authors:

 We added the following sentence: “However, this relationship can reverse when using very high spatial resolution imagery (Arenas Castro et al. 2020).” [Note: the author-date citation here is for clarity; in the revised manuscript it is a numbered citation.]

 

- Line 375. It does not have to be if the atmospheric correction has been done well.

Answer of the authors:

 We compensated the cloud contaminated pixels with another images. We are pointing to the same small number of cloud contaminated pixels that are left at the edge of the cropped area of image.

We rephrased the sentence accordingly: “Furthermore, residual yet marginal cloud effects coverage on the image from March 07, 2017 could have influenced the performance of the Landsat 8 scene.”

 

- Line 377-379. The authors could confirm, and therefore reinforce, these statements based on the response curves or variable importance derived from the ENMs.

Answer of the authors:

Appendix I Table 2 shows the variable importance. Precipitation variables are the second and third most important variables behind solar radiation. We added the following sentence: “This finding is also reflected by the variable importance of precipitation in the ENMs (Appendix I, Table 2).”

 

- Line 380. Diverse in what way?

Answer of the authors:

 “Diverse” has been replaced by “species-rich”.

 

- Lines 385-386. Hard to understand by broader audience. Try to provide references. E.g., [2].

Answer of the authors:

These lines now read “Such an increasing saturation of species richness could enhance the functioning of ecosystems (Mateo et al. 2017). However, individual alien species may also modify important ecosystem functions towards negative effects even centuries after their establishment when replacing other key species such as dominant plant functional types (Kumar Rai & Singh 2020).” [Note: the author-date citations here are for clarity; in the revised manuscript they are numbered citations.]

 

- Lines 386-388. Any reference to support this statement?

Answer of the authors:

 We refined this statement as follows and supported it with references: “The replacement of one dominating plant functional type by another can particularly affect sensitive ecosystems on very steep slopes in a humid zone. The natural stability of the laurel forest on these slopes is astounding and a result of its species diversity and the clonal root systems of the contributing tree species in combination with their evergreen foliage [44]. In consequence, such a regime shift away from long-lived clonal evergreen trees can create new risks for the human population downslope through altered run-off and erosion potential. The respective loss of diversity caused by an invading species is impacting ecosystem stability in addition [70].”

 

- Lines 400-401. Please consider this statement and support with a deeper search on related bibliography.

Answer of the authors:

We refer to Kumar Rai & Singh (2020) for a further read about the effects of invasive species on biodiversity, ecosystem functioning and services.

 

- Line 407. “…Only in valley bottoms two rare deciduous trees can be found naturally. …” This sentence is hard to locate in the narrative. Please rewrite or be more explicit.

 Answer of the authors:

The message was that there are no deciduous trees in the native ecosystems but only two extremely rare native species (Salix pedicellata subsp. canariensis (C.Sm. ex Link) A.K.Skvortsov; Sambucus palmensis Link) along the semi-permanent brooks and streams. We rewrote this section and illustrated the statement more explicitly. Such native deciduous trees can hardly be detected by the sensors we used not only because they are very rare in nature with just a few specimens on the entire island which is known because of their high value in nature conservation, but also because these specimen are located at the bottom of deep valleys in strongly shaded ravines.

 

Figure 1. To improve the self-explanation of figures, please provide different dates for each image (boxes).

Answer of the authors:

We agree that the provision of the dates for each satellite image is necessary and essential for the reproducibility of our analysis. The dates are provided in a separate table for better overview and we think that the repetition of the dates in the figure captions would distract from the main messages of the figures and decrease their intuitive understanding.

 

Figure 2. Please change the color ramp for changed pixels because there is not a good contrast with the background.

Answer of the authors:

We prepared this and other maps again with a changed colour scheme in order to improve the contrast.

 

Figure A (SM). Please change color ramp.

Answer of the authors:

 We changed the color ramp for a better understanding.

 

Table 2. How and why these thresholds were choosing? In addition, could the area be the other way around?

 Answer of the authors:

 The thresholds were chosen based on the maximization of sensitivity and specificity and are a common measure to translate habitat suitability maps into binary maps. As for every model the threshold must be calculated individually, these differ slightly between the maps based on remote sensing detected trees or field observations.

Concerning the area, we scrutinized the calculations again and switched the numbers. You saw this correctly and the numbers were initially written the wrong way around.

 

Additional references suggested in this review.

 Answer of the authors:

We integrated these and other new references to the paper. Also, we changed the in-text citations to match this journal’s citation style. Please see the extensive and tracked changes across the revised manuscript.

 

[1] Manel, S., Williams, H.C. and Ormerod, S. (2001), Evaluating presence–absence models in ecology: the need to account for prevalence. Journal of Applied Ecology, 38: 921-931. doi:10.1046/j.1365-2664.2001.00647.x

[2] Mateo RG, Mokany K, Guisan A. Biodiversity Models: What If Unsaturation Is the Rule? Trends Ecol Evol. 2017;32(8):556-566. doi:10.1016/j.tree.2017.05.003

[3] Arenas-Castro, S.; Fernández-Haeger, J.; Jordano-Barbudo, D. Evaluation and Comparison of QuickBird and ADS40-SH52 Multispectral Imagery for Mapping Iberian Wild Pear Trees (Pyrus bourgaeana, Decne) in a Mediterranean Mixed Forest. Forests 2014, 5, 1304-1330.

[4] Andrew, M.E. and Ustin, S.L. (2009), Habitat suitability modelling of an invasive plant with advanced remote sensing data. Diversity and Distributions, 15: 627-640. doi:10.1111/j.1472-4642.2009.00568.x

[5] Bolch E.A. et al. (2020) Remote Detection of Invasive Alien Species. In: Cavender-Bares J., Gamon J., Townsend P. (eds) Remote Sensing of Plant Biodiversity. Springer, Cham. https://doi.org/10.1007/978-3-030-33157-3_12

[6] Huang CY, Asner GP. Applications of remote sensing to alien invasive plant studies. Sensors (Basel). 2009;9(6):4869-89. doi: 10.3390/s90604869. Epub 2009 Jun 19. PMID: 22408558; PMCID: PMC3291943.

[7] Rocchini D, Andreo V, Förster M, et al. Potential of remote sensing to predict species invasions: A modelling perspective. Progress in Physical Geography: Earth and Environment. 2015;39(3):283-309. doi:10.1177/0309133315574659

[8] Vaz AS, Alcaraz-Segura D, Vicente JR and Honrado JP (2019) The Many Roles of Remote Sensing in Invasion Science. Front. Ecol. Evol. 7:370. doi: 10.3389/fevo.2019.00370

 

 

 

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Thanks for addressing my concerns.

Author Response

We did a spell check throughout the entire manuscript, which was done by Prof. Richard Field from Nottingham.

 

Reviewer 4 Report

Dear Authors,

 

I appreciate the effort to improve the manuscript. Most of the issues that were raised in the first review round were satisfactorily addressed. Thus, I think the paper is almost ready to be published in Remote Sensing after some minor issues that are needed to clarify. Please check my comments below.

 

General comments

The introduction was reinforced in terms of related references and allows to give greater robustness to the state of the art. The objectives of the study are also clearer. The order of the sections and their headings in Methods and Results, as well as tables and figures, now allows a greater lightness in the reading that makes it more coherent, in addition to favouring its replicability. And the discussion has also been reinforced with specific references to other case studies. However, and despite the efforts made, I still find gaps in the methodology that can probably be resolved by rewriting the text. This is the case of the application of ENMs through ensemble modelling approaches such as biomod2.

Apparently, the authors applied ensemble models to predict the habitat suitability of the chestnuts. Despite providing the results in Table I (Appendix I), it is not clear in the text how they performed the ensemble approach. Normally, the prediction of the ensemble model made from the "combination" of the individual models is usually showed. However, it is still not clear whether the habitat suitability maps correspond to ensemble models for each individual algorithm, which does not make much sense, or to the ensemble from the all individual models, which is the final purpose of biomod2 (see specific comments below). From the point of view of the replicability of the study, it is advisable that these details are well explained and clear. I recommend being more explicit in order to describe which are the results derived from the individual models, and which from the ensemble approach.

 

Specific comments:

- Line 230. Why then use image from March? It is not clear if the authors used a combination of images between periods, or just individual images. It was used just Sentinel for this purpose? Please explain it.

- Line 259. This refers to Table C in SM? Because data provided do not match to 70%/30%. Please explain it.

- Line 295-297. Please trying to better explain how this was done.

- Line 307. “… 300 m and 100, respectively …”, but based on which criteria? Please explain it.

- Line 308. Why four? They are very few rounds. Maybe to make the computing time profitable?

- Lines 319-321. Ensemble models considering what, for individual models? Normally, the ensemble is made with the predictions of individual algorithms considered all together and considering a threshold that is not clearly explained here. Please be more explicit. In addition, why 50 and 34 occurrence data for that purpose? Please explain it.

- Line 472. This statement does not appear in this table. In fact, there is no table in which the contribution of the variables is shown.

- Line 580-582. Please check the numbering of the references in the text so that they match in the bibliography. E.g., [67] and [76].

 

Table C. Perhaps remove “… result validation …”?

Author Response

Response to Reviewer 4

 

 

Dear Reviewer 4

 

We highly appreciate your additional remarks and the fact that you consider our paper now close to being ready for publication. In the following we are addressing your latest points and hope that now all critical points will be solved. All co-authors were informed about the additional requests and involved in solving them.

 

Carl Beierkuhnlein

 

 

- Line 230. Why then use image from March? It is not clear if the authors used a combination of images between periods, or just individual images. It was used just Sentinel for this purpose? Please explain it.

 

For each sensor (not only Sentinel), we regressed one summer (chestnut in leaf) image against one early spring (chestnut leafless) image. A slight complication is that February data from another image were substituted for parts of a March image that were obscured by clouds. So one image was a composite, in one case. We have tried to make this clearer via the following text now in the Methods:

“We use image regression with the Landsat 8 images from March 07 and July 29, 2017, and with the Sentinel-2 images from July 08, 2018 and February 13, 2019 (Appendix A). […]

However, the part of the Landsat 8 image from March 2017 that contains cloud was cropped out with the help of Quality Assessment band shipped with the Landsat 8 surface reflectance product and compensated with an image from February 03, 2017 after histogram matching in R using the package RStoolbox [47].”

 

As mentioned, we used image regression to compare the two images for each sensor. Significant changes in reflection, such as those associated with leaf-on vs leaf-off can be identified as large residuals. Thus we identified occurrence of C. sativa, with the help of random forest to eliminate ambiguity with other agents of change.  Because the analysis is of the size of the change (or difference) between the two images, it does not matter which one is chosen to be the first image. We have clarified this by changing what was previously line 230, and the preceding sentences, to read:

“Therefore, an image from one date can be regressed against the image from another date using least-squares regression [39,48].

[…] We considered the image from one date to be a linear function of the image from the other date. Therefore, the image from date 1 was regressed on the image from date 2. We arbitrarily assigned the images from July as date 1.”

 

 

- Line 259. This refers to Table C in SM? Because data provided do not match to 70%/30%. Please explain it.

 

Yes, Table C in SM. From the table we have for Sentinel-2 training data = 101501 and testing data = 43499. This makes total of 145000. Given that 30 % of 145000 = 43500 and 70% of 145000 = 101500, the figures work (they round to 30% and 70%).

 

Similarly, for Landsat 8 training data = 11557 and testing data 4952, total data = 16509 and 30% of 16509 = 4952.7 and 70% of 16509 = 11556.3

 

 

- Line 295-297. Please trying to better explain how this was done.

 

We rewrote this part of the text to try to make this clearer as an overall description of our modelling; the more detailed description comes in the penultimate paragraph of the Methods.  We changed the specific sentence to: “To obtain a habitat suitability map for C. sativa, we applied generalized additive models (GAMs), Maximum Entropy (MaxEnt) and RandomForest (RF), combining them into an ensemble model (EM) using biomod2 [54] (see results for each model algorithm in Appendix F).”

 

- Line 307. “… 300 m and 100, respectively …”, but based on which criteria? Please explain it.

 

We applied 300m for RS data and 100m for field data because the RS data were uniformly sparse but field data were clumped (it was not possible to reach all the places physically so we cannot collect data from all the places, this often happens in the field based study). Applying 300 m in field data would result in far fewer species occurrences. We also took into consideration that the environmental rasters that we used are of 100 m*100 m spatial resolution. As we wanted to avoid more than one species occurrence in a single pixel we used 100m as a thinning parameter.

 

We added the following, in lines 305-310: ‘We used 300 m for RS data thinning and 100 m for field data thinning because the RS data were uniformly rasterised and field data were clumped due to inaccessible field sites. Applying 300 m in field data would result in far fewer species occurrences. The rationale for a 100 m minimum distance is that the environmental raster data that we used has a spatial resolution of 100 m. Hence, we wanted to avoid more than one species occurrence point in a single pixel.’

 

- Line 308. Why four? They are very few rounds. Maybe to make the computing time profitable?

 

In total we have 40 repetitions to build each ensemble model (four times with 10 different pseudo absence sets per algorithm), which is a decent number for replications. The primary reason we did not choose a higher number is, as you mention, computing time.

 

- Lines 319-321. Ensemble models considering what, for individual models? Normally, the ensemble is made with the predictions of individual algorithms considered all together and considering a threshold that is not clearly explained here. Please be more explicit. In addition, why 50 and 34 occurrence data for that purpose? Please explain it.

 

Yes, we used individual models to build our ensemble models in exact the way you describe: based on the predictions of all individual algorithms considered together. This now reads: “For EM projections, only models meeting the quality standards of total true skill statistic (TSS) > 0.7 and area under the ROC curve (AUC) > 0.8 were used. Individual models that did not meet these requirements were excluded from building the EM – including all the GAM and MaxEnt models [Appendix G1 and G2]. Our resulting EMs were based on 50 and 34 single models for RS and field occurrence data, respectively.”

 

That new text also explains why 50 and 34 models (not occurrence data – the numbers of individual models used for the ensemble model, based on remote sensing and field occurrence data) were used. We are sorry that the way we had previously described those numbers in the text was misleading.

 

 

- Line 472. This statement does not appear in this table. In fact, there is no table in which the contribution of the variables is shown.

 

We removed this statement and the reference to the table, which were relics from a former version. Thanks for spotting this mistake.

 

 

- Line 580-582. Please check the numbering of the references in the text so that they match in the bibliography. E.g., [67] and [76].

 

We checked the referenced in and assigned Müllerova et al. to [77]. We checked the entire document and did not find further wrong attributions to references. Again, thanks for spotting this.

 

Table C. Perhaps remove “… result validation …”?”

 

Has been done in the appendix.

 

 

 

 

 

Author Response File: Author Response.docx