Using Unoccupied Aerial Vehicles (UAVs) to Map Seagrass Cover from Sentinel-2 Imagery

Round 1

Reviewer 1 Report

Please address the comments below:

line 100: what about the application of OBIA in heterogenous pixel images? It seems that this approach may be suitable for this scenario. 

lines 111-115. Vague, please rephrase.

line 163: Please, be consistent in naming Random Forest with or without capital letters - compare e.g., line 159.

line 164. Not only canopy cover, but also benthic habitat mapping in various sites and scenarios. Consider supplementing reference: 10.1016/j.scitotenv.2021.149712

line 167: Supplement which software implementation of Random Forest did you use.

line 187: did you mean ArcGIS Pro? Please clarify and provide the version of the software

line 216: be careful with Kappa - see 10.1080/01431161.2011.552923

lines 272-282: I may have not noticed, but did you apply any feature extraction/selection techniques? Additional predictor variables like DEM and its deliverables (slope/aspect/rugosity) can potentially greatly increase the accuracy of your Random forest model.

lines 312-313: be careful, see the similar approach: 10.3390/rs13183681

line 422: suggested reference for structure from motion - 10.3390/jmse6030077

line 435: continuous variables are problematic, however, higher resolution can help to distinguish patterns between seabed habitats

Author Response

Many thanks for the comments, please see the document attached

Author Response File: Author Response.docx

Reviewer 2 Report

Dear authors,

I think you have done a good job. Remote sensing mapping of seagrasses is not an easy task and, to confirm this, there are many examples in the literature with questionable results. In particular, the use of the SPM bands seems very interesting to me and I will try to use it.

In my opinion, the manuscript is very interesting for anyone involved in mapping seagrasses activities and is suitable for publication on Remote Sensing after a few very small revisions:

Pag 5, line 163. What software was used to perform the random forest regression?

Pag 5, line 200. I suggest entering a bibliographic reference for the Global Moran's I index.

Pag 10. Lines 303-307. I suggest a few words regarding the Gini Coefficient.

Author Response

Many thanks for the comments, please see the document attached

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript proposed a three-step approach that combines in situ, aerial, and satellite data to map percentage seagrass cover at Turneffe Atoll, Belize. The use of methodology is interesting and appropriate. However, I have some concerns that should be well addressed before further proceedings.

1. How many sites of the photo survey were conducted? And can authors show the range of seagrass percentage of all photo survey sites? Are these training samples enough for random forest regression?
2. The change of water depth due to tide will affect the reflectance of submerged seagrass. Please show the tidal conditions of the UAV survey and acquisition time of sentinel images and explain tidal effects on percentage prediction of seagrasses.
3. The authors mentioned that transfer errors between the three datasets in the paper. It is very important for the integrity of the manuscript to quantify the transfer errors conveyed by each feature of the random forest regressor. It would make the study more convincing.
4. The spatial distribution of seagrass is affected by many factors, such as topography, hydrodynamics, or terrestrial inputs, and so on. It would be better to explain how these factors affect the spatial distribution of seagrass.

Author Response

Many thanks for the comments, please see the document attached

Author Response File: Author Response.docx

Reviewer 4 Report

Review of the manuscript 

Using Unoccupied Aerial Vehicles (UAVs) to Map Seagrass 2

Cover from Sentinel-2 imagery by Carpenter et al

The manuscript is very well written and addresses clearly the different sections for background, methodology, replicability, results and discussion.

The methods are original and well thought, and this is the kind of paper that i like to read, clear and concise, congratulations to the authors.

Although i understand that the drone orthomosaics are used to train the larger images (sentinel) i think i missed the comparison of the results between the use of drone orthomosaics as a middle-step, and the straight up classification using the field data as training sites... it would be a great addition and would give a more robust manuscript.

Other suggestion is the figure 5, it really doesn't do justice to the results. probably a white background would give the reader a best visual assessment of the results, since the black and masked portions tend to overwhelm the data in green and one doesn't know if the darkest parts are NO DATA or really 100% SG cover.

Author Response

Many thanks for the comments, please see the document attached

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The manuscript can be accepted for publication in present form.

Reviewer 4 Report

Fair enough!... i believe the paper is good to go.