Sensitivity of Optical Satellites to Estimate Windthrow Tree-Mortality in a Central Amazon Forest

Round 1

Reviewer 1 Report

Title

“Sensitivity of optical satellites to estimate windthrow tree-mortality in a Central Amazon forest.”

A nice title. You can remove the full stop from the end of it.

Abstract

The abstract is well-written and descriptive.

1.Introduction

“Optical satellite imagery with medium spatial-resolution (10 m – 39.9 m pixel size)…” I did not know there were such thresholds for spatial resolution. I think you can easily find other classification schemes for grouping optical images.

The intro section seems a little bit poor. The authors focused only on optical satellite imagery. They should also provide other approaches estimating tree mortality events (e.g., LiDAR, SAR, temporal forest plan data). Please include some relevant work conducted with data sources other than optical satellites and assess their usefulness compared to your approach.

2.Material and methods

Fig. 1. Please change “Brazil limits” to “Country boundaries” or “Boundaries of Brazil”.

“Terra-firme” and “campinarana”. Please provide additional information on these type of forest community. As a forester, I am not familiar with these terms… Isn’t there any more common, English term for it?

“600 trees/ha” How do you know the mean number of trees per hectare? Is it coming from forest management plans? Please cite your reference.

“photosynthetic vegetation (GV)”. GV or PV? Or you meant green vegetation? Please correct if it is a mistake.

Please write the scientific names of the tree species in italic.

Fig.1. Is the entire forested land productive (canopy cover is greater than 10%). The forestland (green areas) may be classified into 2 groups, as dense and sparsely covered (degraded) forests.

Line 210. “overlayed” pls correct the word.

Although the MS has a subsection titled “Statistical Analysis”, subsection “2.3. Remote sensing estimates of windthrow tree mortality” includes many statistical descriptor and tests as well. Additionally, I would give a table for classifying the severity of windthrow instead of Lines 247-251.

Line 268. “plataform” correction.

3.Results

In table 1. What is R2KL (Kullback-Leibler coefficient of determination)? I could not see any explanation for it in the Methodology section? What is the difference between the adjusted R2 and R2KL? The authors should clarify this.

Lines 339-341. “undisturbed condition (i.e, old-growth forest)”. Old-growth forest may also be affected by other biotic or abiotic disturbances. How did you know all disturbances within your study area stemming from windthrow? I also wonder is this a managed or unmanaged site? It is vague in the MS.

“Landsat has a longer-term collection of images than Sentinel 2 or WorldView 2 [96], which makes it more suitable for landscape-scale studies aiming at mapping windthrows and monitoring forest recovery over time [2,4]. However, Landsat may be inaccurate for quantifying and describing disturbance created by clusters of less than 6-8 fallen trees [16].” This is the take-home message for me. Perhaps, it could be highlighted in conclusion section, too.

4.Conclusion

Landsat imagery is being promoted so much here. It seems a little bit conflicted with Lines 497-498. The statement may be ba

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This study proposed a measurement method using satellite imagery to assess windthrow tree-mortality in the Amazon rainforest. The Spectral Mixture Analysis (SMA) method was used to quantify the fractions of endmembers (photosynthetic vegetation (GV), non-photosynthetic vegetation (NPV), and shade (SHD)).

The results shows that all three satellites (Landsat 8, Sentinel 2, and WorldView 2) produced similar and reliable estimates, but Landsat 8 was most accurate in capturing field observations of variations in tree mortality across the disturbance gradient. This study validates the reliability of Landsat imagery for assessing patterns of tree mortality in dense and heterogeneous tropical forests.

This manuscript is well structured and written. The topic fits well with the scope of the RS. To further progress, I have a small comment for you to include:

Minor concerns:

1. In the Introduction section, the author first introduce the importance and significance of mapping windthrow tree-mortality. Thus, L91-96 should be moved to the first paragraph, as it also introduces the significance of mapping windthrow tree-mortality.

2. L54-L60 introduced the application of optical images in monitoring windthrow tree-mortality. They should be moved to the second paragraph or combined with the third paragraph.

3. In the Discussion section, why the higher resolution images (WorldView) did not better identify windthrow tree-mortality was discussed and explained. I suggest summarizing these findings in the Abstract section of the article so that readers can quickly understand the main points of the study.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper uses a combination of field data and optical data from three satellite constellations at different resolutions to examine and model tree mortality due to windthrow in an area of the Amazon forest. The modelling approach is based around GLMs to model mortality based on changes in an SMA of the satellite data before and after the windthrow event.

Overall the paper is a useful contribution, and it is important to see the field data being used to verify what is observable accurately from the remote sensing products, so I is good to see publications trying to maximise use of the ground-truthing data appropriately. Overall the paper is well written - though the statistical analysis is a little rushed over. 

There are some points in the more detailed comments below for longer discussion and comparison. The two major points that I think need a bit more discussion are around what would happen for the high resolution optical data if it was downgraded to a lower resolution? That would seem easy enough to do and would demonstrate that it really is resolution that is key. Moreover it would be good to discuss why that is the case a bit more clearly - is it just because of the field data being used corresponding to that resolution better. Lastly I think figure 3b shows that at low mortality the data seems to show almost no correlation between satellite and field data mortality estimates. I think that warrants a further discussion as it doesn't seem that the description in the paper really acknowledges this, and commensurately the conclusions drawn from the analysis then feel overstated.

General comment:

Why use the statistical approach given rather than a machine learning one. Particularly as there is already some success with use of the latter eg Automatic Windthrow Detection Using Very-High-Resolution Satellite Imagery and Deep Learning Dmitry E. Kislov and Kirill A. Korznikov Remote Sensing 2020, 12, 1145. Some of that approach should be discussed, and that paper cited somewhere in the current paper. 

Details in the main paper: 

L176 - The selection of 'endmembers' is done by hand. Surely a better way to proceed would be precisely to use the ground data to build a machine learning model to detect those pixels, rather than the entire approach outlined here which is more user intensive? 

L186 - As I understand equation 1 the SHD fraction is effectively being discounted. I think it is necessary therefore to comment on how that might change across satellites and temporally across acquisitions. Presumably it is quite variable and so this is something fairly uncontrollable in the analysis. Likewise would a systematic change to SHD not be likely in an windthrow event? It would be good to add comment on this.

L255 - The tests for normality and homogeneity for the field measurements seem a bit confusing - there has been no explanation of why we might expect either of those results on field plots of the size recorded? I think it needs either a justification for why this would be worth looking for or expected or else to be removed as the later sentence seems to indicate the plots don't exhibit a normal distribution in fact and a nonparametric test is used.

L262 - Why is a non-paired test used? Surely for the field plots with two methods of assessment of mortality (field and remote sensing), the situation is paired - am I misunderstanding how the test is being applied?

L272 - I think it warrants a discussion here about the sampling methodology used to select the different plots. Particularly thinking of scaling over a region the means of sampling is a key issue in the validity of any result here about tree mortality numbers and how that changes across a region (particularly as the satellite sources of data sample a wider nearby region).

L282 - I think the resolution needs some discussion here, when comparing the different satellites. Is the effect coming from resolution alone - could be easily ascertained by downgrading the higher resolution satellite products to the same resolution as Landsat using eg a mean or median filter applied to downsample the pixel values. That could also be done on raw spectral information before the SMA, or by aggregating and downgrading the individual spectral bands that make up the mixture GV/NPD/SHD. Would be good to get a sense of this as it seems to me likely that some of the benefits here of Landsat are effectively overstated and simply come from an aggregation of data to a coarser scale (presumably a scale where SMA works a bit better or correspondence to the field plot size is better).

L284 - Whilst I agree that the evidence presented shows increasing the resolution of the examined optical data lowers agreement with the field data I think this warrants more discussion as it is not as simple as this paragraph makes the analysis sound.

- Is the difference due to resolution or changing the satellite product, these are confounded in the current analysis, the above suggestion about downgrading the higher resolution products would clarify this, and I strongly suspect it is really just resolution not satellite that makes a difference. If so the satellites should be removed from this description and a reference made to resolution considered instead.

- The resolution of the ground data is relatively coarse (not a criticism, it is clearly hard to collect). But could the result be different if this was supplied at a higher resolution, i.e. is the agreement with Landsat just because that matches up the training data resolution best? Would be good to comment and discuss.

- Lastly the intercept of the model seems odd to me, if Delta NPV is zero and that is our predictor variable if I understand correctly; surely we would expect no wind throw damage in that case. Is this because the model is only/largely trained on plots where some windthrow is observed? In which case the intercept area of the model might not be reliable?

L324 - Looking at Fig 3b if I were to separate off <40% mortality as predicted by satellite. It seems to me that the left side of each of those plots has almost zero correlation between the field data and the satellite. The trend is only recovered when the high mortality cases are mixed in. This seems to me to be a major issue with the modelling the ability to distinguish cases where there are low mortality rates is very poor indeed. This needs further comment and discussion as it is not very clearly stated in the manuscript at present. Could the different cases be separated out into different models (i.e modelling within a band of windthrow mortalities)?

L364 - It isn't clear to me that this statement is supported by the data shown here. More demonstrative of that statement would be a comparison of the SMA split for each satellite and a demonstration that over time in a greening and damaged patch the GV/NPD split changes faster for the higher resolution products.

L372 - again not entirely demonstrated by what has been shown here.

L377 - again here I think resolution is more interesting as a way of describing this shift rather than a focus on particular satellites. 

L393/394 - This seems to backup my suggestion above that some of the result observed in the paper is due to the correspondence between the field scale measurement and Landsat resolution.

L421 - Given the possibility of geolocation errors it seems even more important to state something about survey design and methodology for selection of sites, since implicitly a set of sites is effectively representing a larger region.

L433 - This doesn't seem that robust as a metric in the low mortality case according to the plot shown in figure 3b.

In the appendix:

- The methodology text at the beginning of appendix A concerning the method for estimating the number of dead trees per hectare in the field plots is significant enough that it should be included somewhere in the main paper. 

- Figure s9 has typos in the x-axis labels.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Many thanks to the authors for considered responses to comments.

Replies to responses below:

Response 2 - My comment was really raising that a separate strand of work using ML exists to address this problem, and as such should therefore be referenced and discussed in the introduction. This is not present in the updated draft.

Response 3 - If there are no clear end-members as the response seems to suggest, with pixels generally more diffuse, then that throws into question all of the SMA process surely?  Anyway the initial comment is addressed.

Response 4 - Thankyou for confirming that the SHD fraction is excluded from what follows and the other fractions are assumed to sum to 1. The issue of systematic bias introduced as changes occur to that fraction is therefore relevant - the text should mention this more clearly as a potential issue with this sort of analysis. To be clear I think you have mitigated as much as is possible, but there remains the possibility of introduced bias in this way during across an event, or across satellites.

Response 5 - I think all that is needed is to be clearer in the text that the tests are to determine which are the appropriate statistical tools, that wasn't obvious to me sorry.

Response 7 - I'm glad this has been better addressed in the manuscript, but the question of representative statistical samples is a key one and so it is important that some discussion of the potential for these sites not to be representative of a wider region is discussed in the main paper not just in the supplementary material.

Response 9 - I was commenting the complete opposite, that I think downgraded resolution from one of the high resolution satellites would answer many questions and allow you to state your inference much more clearly. I think this is clear from the material included now in the supplement.

Response 10 - I appreciate the elements added to the discussion. The central issue remains however that the fits in 3b are just not good, clearly from the scatter of points it makes little sense to fit one straight line through the data - as ultimately evidenced by the low R^2 values. Under 40% on the plots just has no clear correlation with the field data at all - I think that should still be more strongly acknowledged in the text.

Line 821 - shall -> should 

Responses 11-18 - Thanks for your efforts in changing the manuscript. I think the new version reads much better thank you for clarifying the text.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf