the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Synergistic use of Sentinel-2 and UAV-derived data for Plant Community Cover distribution mapping of coastal meadows with Digital Elevation Models
Miguel Villoslada Peciña
Raymond D. Ward
Thaisa F. Bergamo
Chris B. Joyce
Kalev Sepp
Abstract. Coastal wetlands provide a range of ecosystem services, yet are currently under threat from global change impacts. Thus, monitoring and assessment is vital for evaluating their status, extent and distribution. Remote sensing provides an excellent tool for evaluating coastal ecosystems, whether with small scale studies using drones or national/regional/global scale studies using satellite derived data. This study used a fine-scale plant community classification of coastal meadows in Estonia derived from a multispectral camera on board Unoccupied Aerial Vehicles (UAV) to calculate the Plant Fractional Cover (PFC) in Sentinel-2 MultiSpectral Instrument sensor (MSI) grids. A Random Forest algorithm was trained and tested with vegetation indices (VI) calculated from the spectral bands extracted from the MSI sensor to predict the PFC. Additional RF models were trained and tested after adding a Digital Elevation Model (DEM). After comparing the models, results show that using DEM with VI can increase the prediction accuracy of PFC up to two times (R2 58–70 %). This suggests the use of ancillary data such as DEM to improve the prediction of empirical machine learning models, providing an appropriate approach to upscale local studies to wider areas for management and conservation purposes.
- Preprint
(8252 KB) - Metadata XML
- BibTeX
- EndNote
Ricardo Martinez Prentice et al.
Status: open (until 15 Oct 2023)
-
RC1: 'Comment on bg-2023-95', Anonymous Referee #1, 24 Aug 2023
reply
This paper talked about the upscaling of Drone image classification to open access satellite images, and applied to the coastal wetlands. I do think it is a good idea to overcome the disadvantages of low spatial resolution of open-access satellite images. It is a key step to achieve the continuous monitoring of coastal wetlands at a low cost.
However, I have a few concerns about this method. I suggest this manuscript needs to be major revised.
If I understand correctly, the author establishes an individual RF regression model for each community to each study site. My concern first concern is that have the authors united the plant fractional cover (PFC) of each landcover to make sure the sum value of the PFC of each community within each Sentinel pixel is equal to 1. If so, please highlight it or remind me where I can find it in the main text. If not, I suggest the authors to rescale the retrieved results and test the accuracy again. Some previous analyses have indicated that such rescaling can change the accuracy obviously (e.g., Immitzer et al., Remote Sensing of Environment, doi: 10.1016/j.rse.2017.09.031; Yang et al., Remote Sensing, doi: 10.3390/rs12193224).
Here are some specific comments in the main text:
- Do the authors think that tidal level would affect the accuracy of PFC estimation? Some previous works mentioned that tidal level can significantly affect the land cover classification (e.g., Kearney et al., 2009 Journal of Coastal Research, doi: 10.2112/08-1080.1). I suggest that the tidal level at the time of each drone image and Sentinal image acquisition should be reported.
- The authors examine the accuracy of applications with and without elevation data provided by DEM. Can authors specify that did you use the same points to train and test your two applications? I think it is important to show that adding DEM is useful. In addition, can the authors show the plots of DEM in the main text or supporting information?
- The introduction section needs to be reshaped to introduce the topic step by step. For example, you talked about remote sensing in line 20, and talk about wetlands, and then move to remote sensing again. Another example is that Lines 33- 40 are likely to appear in the study areas section.
- Line 39. There are two references, i.e., Ward et al., 2016a and Ward et al., 2016. I did not find Ward et al., 2016a in your reference list. Another obvious error is Line 370. So I suggest the author please check this and also other references.
- Line 70. Please define VI, although you have defined it in the abstract. In addition, there are too many abbreviations, some of which were just used a few times, making the manuscript difficult to read and understand. So please remove unuseful abbreviations. And I also suggest the authors construct a table to explain each abbreviation.
- Line 90 - 95. Please specify the manufacture of your drone here. Please also specify the procedure of your radiometric correction, parameters and models used here. I think they are also useful to other researchers to do similar things.
- Line 95 – 100. Can you also please show the confusion matrix for your classifications here? This would help to show that your selection of RF makes more sense.
- Line 123. Can you please explain more about the accuracy of the DEM used here. From my point of view, the accuracy of lidar-dem over wetlands is a bit low. So I think the specification is useful to show the robustness of your method.
- Table 2. I am not sure I lost something. But I do not know what @ means in this table. Please explain it in the table caption. And please show the unit (probably nm) of each band.
- I think the first row of each table can be highlighted, making it easier to read and review.
- I also suggest that Figs. A1 and A2 can appear in the main text to better display the accuracy.
- In the discussion part, I suggest the authors explain the value of the proposed method in future research.
Citation: https://doi.org/10.5194/bg-2023-95-RC1 -
AC1: 'Reply on RC1', Ricardo Martinez Prentice, 18 Sep 2023
reply
Here are our answers in a document (pdf format). Thank you very much for your comments.
-
RC2: 'Comment on bg-2023-95', Anonymous Referee #2, 23 Sep 2023
reply
In this paper the authors use a Random Forest model to derive the Plant Fractional Cover by upscaling UAV multispectral data to Sentinel-2 MSI data. They present a method that allows to overcome the limitations of low-resolution images from satellite data by using UAV. The topic fits the purpose of Biogeosciences and the special issue. However, I believe the manuscript needs to go through several revisions as I have comments and questions to the authors regarding the methods, the results, and the overall presentation of the study.
TITLE: Here you use the Plant Community Cover term but in the paper the Plant Community Fraction is used. I suggest correcting that.
INTRODUCTION: the authors provided a nice introduction to the topic of the paper. They provide information about the advantages of using remote sensing to monitor wetlands, pros and cons of using UAV and Satellite-based imagery, and the potential in the upscaling of UAV images to satellite resolutions. However, the introduction does not clearly state what is the novelty of the study. From my understanding of the last paragraphs, the implementation of Machine Learning (ML) models and especially Random Forest (RF) models to infer Plant Fractional Cover (PFC) is not new. At the same time it is not new to couple these models with DEMs as ancillary data. The last sentences suggest that this is what the study aims to. I think the introduction need to point out what are the new aspects the authors are looking at. What does this paper add to the current state of the field? Was sentinel-2 never used? Is the scale of the study new?
I also fell like the authors could reorganize a bit the structure of the introduction. At the beginning you mention remote sensing, then move to wetlands, and then back to remote sensing. I would first introduce the wetlands and then the remote sensing aspect. It makes the reading more fluid.
LINE 39: There is a Ward et al. (2016) and Ward et al. (2016a) cited but only one referenced in the bibliography. Is it a typo? Please correct or add the citation to the list.
LINE 71: Acronym VI was not introduced before.
LINE 75: What’s the tidal range here? Do these areas go underwater regularly at each tidal cycle? During high tide/spring tide? Or only rarely during storms?
LINE 76: Here you use Figure 1 to reference to the figure. Many times throughout the paper you use the short form Fig. to reference to figures. I suggest picking one style and being consistent.
LINE 82: I have to say that I am not familiar with the plant species here considered. I think it would be important to provide more information on these species/communities. For instance, what is their phenology? This would be a very important information. If they have a growing season in a specific time window, it means there is a limited period of the year of useful remote sensing images.
FIGURE 1: This is just a personal preference a very subjective suggestion: I would increase the font size, especially for the latitude and longitude coordinates. I would also identify the study sites not with numbers but with their names and acronyms directly in the map instead of giving a legend in the caption. It would be much easier to locate them.
LINE 96: Would it be possible to provide the confusion matrix for the classification. It is the best way for the reader to quickly assess the quality of the classification. You could simply put it in the Appendix.
LINE 101: I understand that the classification is based on a different study. However, since it is key to the analysis, the authors should provide a few more details on the classifications (see comment above).
LINE 107: The analysis used only one Sentinel-2 scene. Do the authors think that the study would benefit from using multiple scenes? Since the authors say that the method has implications on monitoring (i.e. using multiple images), I think that testing the model on a different scene would prove that the method is more robust. Having said that, I am not suggesting that the acceptance of the paper should depend on this additional analysis as I believe it is a serious task to undertake.
Was the Sentinel scene taken during low tide? I think the presence of water can affect the reflectance. The authors should specify (maybe in the discussion) what kind of images are good for this method. In the discussion there is only the mention to take UAV images close to the satellite passage. The authors can expand here on what are other important aspects that goes into the choice of the scenes.
TABLE 1: I think Table 1 is missing some information. I see only the study area and the drone flight dates, but no tile number or satellite overpass date is reported. Either reformulate the caption or add the information in the table.
TABLE 2: units of wavelength and spatial resolution are missing.
LINE 122: could the authors provide the vertical error of the DEM? Since the microtopography is important, it would be useful to know what’s the vertical error.
FIGURE 2: Like for Figure 1, I suggest to directly use site acronyms in the map. I also suggest increasing the size of the scale bar. It’s very difficult to read it.
LINE 137: ‘A correlation and a linear function were used’ is not very precise. From this I understand that there are two levels of evaluation. What kind of correlation? Is the correlation found with a linear function? Please rephrase for more clarity.
LINE 138: Does the averaging of the elevation smooth the microtopography? Do you think this step has an effect on the performance of the model?
LINE 140: I am a bit confused on the separation of DF1 and DF2. DF0 is a sub-sampled dataframe from DI, and it already contains the elevation variable since the DEM information was previously added. Thus, I understand that DF2=DF0, and it is not really a new dataframe. Would it be easier to consider just two dataframe? One with and one without elevation?
LINE 155: I think the reference to Figure 1 is wrong. Please check.
LINE 169: I assume that in each MSI pixel we want to have the sum of all PFCs equals to 1. If you have a separate model for each plant community, how do you make sure of that? Do you force it somehow? Please clarify this as I think is an important step.
LINE 173-174: I am not sure the reference to Figure 4 is correct: it looks more correct to reference these sentences to Figure 5.
LINE 174: rewrite as ‘predefined hyperparameters’
RESULTS: Here you clearly show that elevation is a key feature to predict vegetation zonation. That makes 100% sense. I think it would be important to show the DEM to the reader due to the importance of this parameter. That would help to understand whether a species prefers a high or low area, which is directly linked to the ability to withstand a low or high hydroperiod.
LINE 185: it would be good to add a figure were you show the correlations between the MSI and PS reflectance in the MSI GRID. Maybe this figure can be shown in the Appendix/Supplementary Material.
LINE 194: I invite the authors to consider moving Figures A1 to A4 to the main text, since they show the goodness of the models. Maybe Figure 2 can be moved to the Appendix to avoid overloading the main text with figures. It just shows the grid, so it is not as useful as the other 4 figures.
TABLE 5: It would be better to specify the p-value instead of simply indicate <0.05.
LINE 216: I would be more specific. Are your results comparable to those studies you mention? I think it would be better to expand here the discussion and show a comparison with other similar studies. It would be very informative to know what you did better. Did these studies consider DEM or only VIs?
LINES 210-224: I think this part is not really useful. Here you are just repeating the methods and giving more results. You can move lines 222-224 in the results sections.
LINE 235: ‘figures 6 and A1’ should be ‘Figures’.
LINE 236: Typo metre. Please correct. Also when you use the number use the unit. So in this case it should be ‘1 m’. Please make sure to follow that throughout the paper.
LINE 237: Are you sure that the reference to Figure A2 is correct? I think the correct reference is Figure A3.
LINE 238: Looking very quickly at Figure A4, elevation seems to be the most important by far with the only exception for the OP class. If I read correctly for the other classes, elevation in terms of importance it is around 0.5, and 3-6 times more important than the VI with highest importance. One could argue that acceptable results could be achieved with only elevation. Have you tried to do that? Can you comment on this?
LINE 239: This comment concerns the entire manuscript. I had hard time to remember all acronyms. I know that they make the writing faster, however I think it would be better to reduce the number. Maybe the authors can simply pick the most used ones.
LINE 242: I think the authors could expand here. Why these communities are so dependent on elevation?
LINE 245 to 248: The sentence starting with ‘Overall, …’ is badly written. I suggest reformulating and breaking it into short sentences.
LINE 251: Do the authors think that other satellite data would have been applicable to the study (e.g. Landsat)?
LINE 257: Can the authors suggest other ancillary data for future research besides DEM? Maybe inundation time?
APPENDIX: The axis ticks are small and hard to read. Especially Figures A3 and A4. In Figures A1 and A2 you are not showing the predicted error. You are just comparing predicted PFC with measured PFC. When you are comparing modelled and observed values you do not need to compute the R2 since you are not really looking for a model between the two. The RMSE values is a good index to evaluate the goodness of your predictions. Maybe you could add a second index like Model Efficiency or Percentage Bias (it doesn’t necessarily have to be these ones). Especially with the second one you could quantify the general tendency of your model to underestimate, and overestimate observed values.
What is the unit/variable in y axis in Figure A3 and A4? Is it the explained variance for each variable? Please clarify.
Citation: https://doi.org/10.5194/bg-2023-95-RC2
Ricardo Martinez Prentice et al.
Ricardo Martinez Prentice et al.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 179 | 42 | 11 | 232 | 7 | 2 |
- HTML: 179
- PDF: 42
- XML: 11
- Total: 232
- BibTeX: 7
- EndNote: 2
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1