Detecting Deforestation Using Logistic Analysis and Sentinel-1 Multitemporal Backscatter Data

Round 1

Reviewer 1 Report

A useful contribution adding a new technique for detecting deforestation using time series of SAR data from Sentinel 1 without needing training data of deforestation instances (which are more difficult to obtain). The paper is nicely presented, and should be published in my view.

A couple of general points:

It would be of interest to explore how effective the proposed approach is for forest degradation as opposed to clearcut, and just make some caveats in the presentation of the approach about how well it would perform in such situations.

How does the proposed approach differ from a logistic regression on a suitable feature space of time series to find deforested pixels - it must be very nearly the same?

It might be helpful to justify the use of the logistic curve inflection point - it makes sense as being the 'centre' of the curve, but also seems to add a bias to the time of deforestation - maybe there is something there to discuss. For instance there is a question that once the change is found maybe the deforestation event should be back-dated to the point where the curve begins to change, or shows a 10% change etc. This would also reduce the bias perhaps? What is the typical timescale of the 'width' of the logistic function transition? Maybe worth a little more discussion of the issue in the paper.

Some more detailed comments:

Section 2.2 - this seems something of a flaw in judging the effectiveness of change detection against the forest mask generated by hand (the effectiveness of by-hand change detection is unclear from the stated method). What is done here about cloud cover missing data - does this not alter the noted deforestation event dates? Why not use one of the large global forest masks such as the Hansen data set where the associated errors (though perhaps more significant when examined at local level) are at least understood. I think this section warrants some more discussion of limitations and justification.

Figure 3 - does this not depend quite critically on the choice of forest/non-forest pixels though?

Figure 4 - shows a very clear example where the method works, it might be helpful to add some extra panes to this plot to show some others where the method is less clear to give a sense of how effective it is and how it might fail for the mis-classified changes.

line 244 - Is there any justification for the choice of logistic function over any other form for the transition function here?

line 277 - again if the drop to linear-like behaviour is undesirable maybe this would motivate thinking of use of a different functional form? (error function say).

line 310-323 - I don't think it is really needed to define precision, recall and f1 here, they are well known.

line 364- 375 - Not sure that it is really shown that the method can predict fine details, as usually fine spatial degradation of forest might be of more interest, and I think that effect is largely uninvestigated here. This whole section really just suggests that a method which takes in some spatial neighbourhood might be more effective at reducing spatial noise across deforested plots, that might be commented on.

line 503 - I think you have misunderstood some of the argument in [22] perhaps. Similarly there the authors don't need deforestation training data, the current study and that one both need some criteria for knowing which pixels are forest initially in order to track changes in those pixels - the difference between approaches is not as significant as the text here suggests.

line 506 - Simply tracking high backscatter pixels isn't necessarily the best approach in all areas - that will not scale well to situations where such high backscatter might be from other sources. That warrants some more caveats I think. Again it isn't clear why a global forest mask is avoided for use to construct baseline forest pixels.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Overall Impression
In this manuscript, the authors introduce a change detection method based on Sentinel-1 SAR backscatter analysis through a logistic regression approach. The algorithm is notable for not requiring training data for purposes of classifying clearcuts and for being able to overcome pixel contamination/omission typical for optical-wavelength images. The algorithm identifies stable forested pixels through a low backscatter variability over the timeframe, then fits logistic curves to the remaining forested pixels to identify breakpoints. The algorithm performs well, both spatially and temporally, based on a simple analysis, and appears to have accuracy comparable to typical dense multitemporal Landsat-based forest disturbance detection methods (e.g., EWMACD, CCDC, BFAST) using optical wavelengths.

While overall the approach is interesting and useful, the manuscript suffers occasionally from lack of consistent definition and rigorous description in the methods. Notation is missing (e.g., Equation 1), or threshold choices are not adequately explained (e.g., an 80th percentile threshold for ruling out which pixels to assess). The authors use a short timeframe in the study, as small as six months in one study area. While this is (eventually) explained, it lends little support to the authors’ commentary regarding making a real-time monitoring system from their algorithm.

If the authors address these and the below comments to make the manuscript more robust, then I would be happy to reconsider it.

General Comments
It would be nice to have some simple regional context maps/insets for the panels in Figure 1, since for example, southeastern Romania is unfamiliar to me (and I expect many readers).

The authors’ study period is only from 2019-2021 for one site, and only March-September 2020 for the other. While this is sufficient to test the algorithm, it precludes any assessment of detection over longer timeframes (in particular, handling multiple disturbances over those longer timeframes).

In section 2.2, the authors refer to the beginning of the monitoring period without stating when that is. Including the years/timeframes here would be helpful. Likewise, the explanation for the short time series in Romania, lines 379-382, would do better in the data section.

In the introduction to section 3, the authors identify cross-polarized backscattering as VH; please give a similar identification for VV to aid the reader.

Equation 1 has variables that are not defined in the text, such as M and Ji. Similarly, the index is not clearly defined. Is i the ith pixel, as suggested by the xy coordinates on the left-hand side?

In the sentence at line 222, the authors propose a percentile-value based approach to thresholding stable forest pixels; this seems arbitrary and risks errors of omission in areas of high forest disturbance. A better explanation of either what data are being percentiled and/or why the 80th percentile is reasonable enough for a general recommendation would help. Alternatively, a discussion of why one simply would not eliminate that threshold and perform logistic regression across the entire map (I assume due to computational cost), or how that might be implemented in cloud systems such as Google Earth Engine, would be appropriate.

There is misnumbering on results subsections (2.x instead of 3.x)

It would be good in the discussion to include paragraphs on 1) the computational costs/drawbacks of pixelwise logistic regression, and 2) how the authors would envision turning this algorithm into a real-time detection method.

Line by Line Comments
31-32: The sentence is strangely past-tense, and the use of “the forest” suggests the authors are thinking of a particular forest instead of forests in general.

211: This sentence is confusing. The authors just stated that river pixels, agricultural areas, and dams show high variability, but then they state that the “standard deviation” (I assume the quantity given in Equation 1, though this is specified as the temporal variability) perfectly separates forest and nonforest pixels. Do the authors mean that this enables forest disturbance detection because the conversion to nonforest shows up clearly?

441: Why use a variability threshold at all? Is it to make the pixel-wise logistic regression more computationally tractable?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

I am generally satisfied by the authors’ responses to my comments, and their revisions accordingly. I can recommend the manuscript be published in Remote Sensing, pending a final editing pass for style and grammar.