Mapping of Eucalyptus in Natura 2000 Areas Using Sentinel 2 Imagery and Artificial Neural Networks
Round 1
Reviewer 1 Report
Dear authors,
Your study presented a data-driven approach to map Eucalyptus tree species by exploiting the capabilities of machine learning techniques, particularly feedforward artificial neural networks, on Sentinel-2 satellite imagery. Specifically, your study first developed a method for automated mapping of Eucalyptus trees using Sentinel-2 imagery, and then applied this method to detect Eucalyptus trees within Natura 2000 areas in Portugal in support of efforts to control invasive species, particularly Eucalyptus trees, outside their designated plantation areas.
It is an interesting and relevant paper. I have some comments and suggestions to improve the manuscript as follows:
- The Introduction is well-written and provides a concise background and overview of the study.
- Under Section 2 (Data and Methods):
- In Section 2.1 (L 80-83), it is not clear how did the authors generated the ground-truth maps. What do the authors mean by ‘created using high resolution imagery from Google Earth’ (for Area 1) or ‘compiled using Google Earth imagery’ (for Areas 2 and 3)? Were Eucalyptus and conifer forests digitised visually? Were vector polygons created to delineate each type of forest? Were annotated training image samples (or tiles of raster images for each type of forest) extracted from the high-resolution imagery?
- In L 87-88, mention the dominant species present in these coniferous forests within the study areas.
- In Section 2.3, L 106-109: need citation/s that have shown FNN architectures perform better than UNet architecture.
- On accuracy assessment, why do the authors report Kappa (Sec 2.4)? Please see the papers by Pontius & Millones (2011) Int J Remote Sens and Foody (2020) Remote Sens Environ, which have recommended the abandonment of the Kappa coefficient for accuracy assessment. The authors should report not only overall accuracy, but also per class accuracies using user’s and producer’s accuracies, including confusion matrices. Please follow the good practice guidance for accuracy assessment of land cover maps by Olofsson et al (2013, 2014) Remote Sens Environ, which complement the other reported statistical measures provided in the study that evaluate model performance and classification accuracy. Also clarify the sampling design used in the study.
- Under Section 3 (Results):
- L 232-235: It would be good to know how many and which Natura 2000 sites have not been significantly affected by encroachment of Eucalyptus trees, which can be presented in a supplementary material. This information can provide readers with a sense of the possible extent or gravity of the encroachment of Eucalyptus trees within all Natura 2000 sites in Portugal.
- L 236-241: These sentences are more appropriate in the Methods section.
- Comments on figures:
- For Figure 5, change the color of the outline of polygons of Eucalyptus plots (for example, to blue instead of black) to provide better contrast compared to the grayscale probabilities and visualisation of the both Eucalyptus and conifer plots.
- Figure 6 can be improved. Include a scale bar and an overlay of the coastline over the probability image to clearly show the extent of the landmass.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Remotesensing-806883-peer-review-v1
Manuscript ID: remotesensing-806883
Type of manuscript: Article
Title: Mapping of Eucalyptus in Natura 2000 areas using high resolution satellite
imagery and artificial neural networks
Comments
This Article was written well and it is important addition to the literature.
The authors used high resolution satellite imagery and advanced machine learning
algorithm such artificial neural networks to mapping plantain forests such as
Eucalyptus in Natura 2000 areas. The result is of interest to readers of remote sensing specialists and remote sensing journals. However, there are several comments that need to be addressed.
The specific comments are listed below:
The abstract needs some revisions regarding the Kappa coefficient in line 9: you don’t
need to use Kappa coefficient in your study because the Kappa is no longer used in the
recent studies relay on my knowledge for many limitations. Recent studies have shown
some limitations of kappa since it gives information that is redundant or misleading for
practical decision making. Previous studies recommend using a more suitable and
simpler approach that focuses on two parameters of disagreement between maps in
terms of the quantity [quantity disagreement] and spatial allocation [allocation
disagreement] of the categories instead of kappa variants. I would advise you to have a
look at the published paper entitled “Death to Kappa: birth of quantity disagreement
and allocation disagreement for accuracy assessment”. Therefore I would greatly
appreciate if you could delete any information related to kappa in the abstract section
and throughout the manuscript. For more details also you can look at the following
paper entitled “Performance of support vector machines and artificial neural network for
mapping endangered tree species using WorldView-2 data in Dukuduku forest, South Africa.
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing Vol. 8, pp. 4825-4840”.
In the lines 6-8 “Feedforward Neura Networks were trained on ground truth maps
compiled using field surveys and high resolution satellite imagery” you mentioned the
training modeling what about the validation data or modeling. Please before your write
information regarding the findings, it is essential to mention how you did training and
validate modeling results?
I would like you to add sentence in end of the abstract section to shows what the study
conclusions (short sentence).
The introduction is mainly well written, however, there are some minor comments:
Line 53; NDVI, please write in full names and then you can include their abbreviations
as you did in the sentence in lines 46-48 (Baccinietal. [20] use data from the moderate
resolution imaging spectroradiometer (MODIS)………………
I would advise you to have a look at the article to write the full names of the
abbreviations and then you can include their abbreviation throughout the manuscript.
In line 58; Since the launch of the Sentinel-2 multi-spectral instruments…………….
Why you are not trying to use other high resolution satellite imagery such as
Worldview or RapidEye, to compare with your satellite (Sentinel-2) for which one could
achieve better results. That mean, it is essential to refer to the important of Sentinel-2
images and their advantages to compare with other high resolution satellite imagery or
what the limitations of the other satellites to justify why you choice the Sentinel-2.
In the abstract you mentioned the objective in this sentence (This study uses multi-spectral
imagery of the Sentinel 2 satellites to map Eucalyptus across Portugal and parts of Spain with a
focus on Natura 2000 areas inside Portugal, that are protected under the European birds and
habitats directives).
and you mentioned in the introduction section (line 65-71); the objectives of your study
(or the main goals) are
1) the development of a method which enables automatic mapping of Eucalyptus trees with
moderate temporal and high spatial resolution (10m), even for small incipient and mixed
populations of Eucalyptus trees, which do not follow the typical NDVI time-series pattern [23]
and 2) the application of this method on Natura 2000 areas in Portugal to detect Eucalyptus
populations and create a data-set that can support invading species control of Eucalyptus trees outside regular plantations.
Your objectives should be the same in the abstract and in the end of the introduction,
but it seems different, please rephrase the objective in the two sections to be
consistency. Is not clear to!
Line 72; move the whole Data and Method section to previous page. There is a gap
Line 142; look at earlier comment, regarding the Kappa coefficient. Any justification?
You typing “Feedforward Neural Networks” in abstract different from the text,
particularly in the discussion section “feed-forward neural network” be consistent,
suggest either Feedforward Neural Networks or feed-forward neural network
In Figure 4; you should type 0 and 1 instead of zero and one. Please look at lines 12,135, 171 and 208.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The intentions of the study are worthwhile. However, I think the paper needs further work before it can be considered for publication. It is not clear where the novelty is, from a remote sensing perspective. If the objective is to present a new technique (i.e., the FNN), then you need to show why it performs better than other modelling approaches. If the objective is to use new data (i.e., Sentinel 2), then you need to show why it is better than other data. Or you could test different combinations of data, or show why adding seasonal information improves results etc. Nothing is really tested from a remote sensing perspective.
If, on the other hand, the objective is to use established techniques in a new environment, then you need more support backing up why you chose this modelling approach, data etc. If the objective is simply to accurately map invasive eucalypts, then I wonder if a different journal with more of an ecology focus is more appropriate.
Neither the introduction or discussion section is very comprehensive. There many many papers that have used remote sensing data to map forest type, tree species, etc. You mention a few, but I think further research of existing literature is warranted. What type of accuracy do other researchers get in two class classifications? I would also not call Sentinel 2 high resolution – it is more medium resolution, especially as most bands are 20m.
In addition, there are over 700 species of eucalyptus, which all have different spectral characteristics. So I am not sure about the practicality of mapping eucalyptus vs non-eucalyptus.
I may be wrong, but it seems that your training data is largely based on plantations (which are easy to map due to homogeneity) but the reason behind the study is more about trying to map where eucalypts have escaped from plantations and are invading natural forests. It may be worth discussing this limitation…
In terms of the model data, it is not clear how many samples were in your training data and how many were in your validation data. And what do these samples represent – are they pixels or patches? What framework were the samples collected in? One would assume a level of randomness has been included in this process?
It is also not clear how you assessed the accuracy of the predicted maps. Sections 2.6 and 2.7 seem to suggest an informal visual assessment. I think this process needs to be more formalised. For example, see Olofsson et al (2014) Good practices for estimating area and assessing accuracy of land change.
A note regarding English, mostly it is OK, but your use of past tense needs some attention. Terms such as ‘have been’ or ‘has been’ should be used for other studies that happened in the past. If you are explaining your methods or what you found it this study ‘was’ or ‘were’ is more appropriate. In several places, I could not tell whether you were talking about other peoples’ work, or your own.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
The paper has been improved since the previous version. A few minor comments as follows:
Line 19: Tasmania is in Australia
In the final paragraph of the Introduction the objective could be more clearly defined
Figure 2 imagery appears quite dark
Line 163: what is the threshold value? I realise it is discussed later, but it seems a little out-of-place here. Perhaps you could say ‘(discussed in detail below)’
Line 258: reword
Figure 9 colours are quite dark. To me, the ‘Area uncertainty’ looks completely black
Line 351: who are the stakeholders?
The manuscript could do with another read and edit for English clarity and style. Future opportunities could be expanded upon in the discussion.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
