Estimating VAIA Windstorm Damaged Forest Area in Italy Using Time Series Sentinel-2 Imagery and Continuous Change Detection Algorithms

Round 1

Reviewer 1 Report

Many typos and misused words. 

Comments provided in PDF. 

Please, justify the processing you did. It seems like much effort to achieve change/nochange results. What does a case where the processing you did is not used and what is the change/nochange accuracy for that? 

The work is interesting. The work is important. The work is needed. And, the work needs more clarification and a thorough read through for English word usage, clarity, and agreement. 

Comments for author File: Comments.pdf

Author Response

Review 1 –

Thanks a lot for your suggestion and revision and the time you spent to give us very poweful comments. We think that the manuscript it is improved a lot following your suggestion.

Please, justify the processing you did. It seems like much effort to achieve change/nochange results. What does a case where the processing you did is not used and what is the change/nochange accuracy for that? 

The work is interesting. The work is important. The work is needed. And, the work needs more clarification and a thorough read through for English word usage, clarity, and agreeme

R: yes our work is intended to demonstrate that Sentinel-2 imagery can be a relevant source of information for mapping the forest areas disturbed by the VAIA storm, investigating the relationship between the time passed after the storm and the accuracy in disturbance mapping. This is relevant for future application in similar occasions to determine when an accurate map of disturbances can be produced on the basis of Sentinel2 acquisitions.

hit the North-Eastern regions of Italy

R: we changed with “hit the northeast” LINE 19 of the Abstract

that destroyed or intensely damaged forest stands spread over an area of 67,000 km2, forested for 144,651.1 ha.

R: we removed the forested area LINE 22

The results hilted that close to the storm

R: we changed the typos word hilted that means highlighted  with “the results showed” LINE 28

Nowadays, remote sensing (RS)

R: we removed nowadays.  LINE 55

Sentinel-1 SAR data [12,17], high resolution satellite imagery [14,18,19], and optical multispectral satellite data at medium resolution

R: we rephrase the sentence with “Sentinel-1 SAR data [12,17], high and medium spatial resolution multispectral satellite imagery” – LINE 57-58

Within the vast availability of different remotely sensed data, the freely available multispectral optical satellite images are still considered the most feasible and cost-effective option to derive information for extensive forest areas [20] since such data are freely available online at medium and fine resolution e.g., such as Landsat (30 m) and Sentinel-2 (S2) (up to 10 m) and with a revisiting time between 16 and 5 days.

R: We rephrase the sentence and we moved part of the text at the end of the introduction followed also the comments of review 2. LINE 68-73 and LINE 118-124

20.000 km²

R: changed with 20,000 km² – LINE 86

underlined that S2 is adequate to detect wind-storm since S2 program offers innovative features for forest remote sensing by combining high spatial resolution (i.e., 13 bands, from 0.443 to 2.190 μm with the visible (i.e., R, G, B,) and the near infrared bands at 10-m spatial resolution and four red-edge bands at 20-m spatial resolution), wide coverage and a quick revisit time (i.e., every 5 days after the launch of Sentinel-2B satellite in 2017).

R: we rephrase and moved the sentences as reported before at the end of the introduction. See LINE 118-124

Reminder

R: change remainder – and we checked all the manuscript LINE 115

i.e., trend, seasonal and reminder) [37,39,42] for each investigated year. Many of these studies use dense Landsat or MODIS TS as input data [37,43], while no ones, for the best of our knowledge, have tested S2 data  or have investigated the possibility map forest areas damaged by windstorms.

R: we rephrase “Many of these studies use dense Landsat or MODIS TS as input data [37,43], while, for the best of our knowledge, no one have tested S2 data and CCDC algorithm to map forest areas damaged by windstorms.” LINE 116-118

of continues

R: we changed the typos with “continuous” LINE 126

a three years S2 Normalized Burn Ratio (NBR)

R: we clarify it added “continually” LINE 130

ground true

R: we changed with truth LINE 134

The study area Coincide

R: we added the s “Coincides” LINE 141

in the following Regions of Italy: Trentino Alto Adige, Veneto, Friuli Venezia Giulia, Lombardia, Piemonte and Valle d’Aosta (Figure 1), for a total of 67,000 km2, that on the basis of locally forest maps available is covered by forest for 144,651.1 ha [51].

R: we rephrased the sentence “in the following Regions of Italy: Trentino Alto Adige, Veneto, Friuli Venezia Giulia, Lombardia, Piemonte and Valle d’Aosta (Figure 1), for a total of 67,000 km2 covered by forest for 144,651.1 ha” LINE 144

many sentences start ths way or end this way. suggest revising “On the basis”

R: We rephrased the sentence “Preliminary field recognition and manual photointerpretation of satellite data showed that a total of 42.525 ha of forests were damaged by the storm [4, 51]” and we cheek through the manuscript the other sentences that start or end in the same way.

Figure 1 - suggest, undamaged

 

R: We changed the legend as you suggest using “undamaged” moreover to be consistent we changed also “damage” with damaged.

suggest using the phrase Surface Reflectance”We used Bottom-Of-Atmosphere S2 images”

R: We added the information about surface reflectance “surface reflectance Bottom-Of-Atmosphere S2 images available”

Figure 2 - what does this figure add to the story in this paper? It is not clear that it provides more information than the study area map at all. Why the red shade?

 

R: The figure represented the monthly cloud free composite of S2 imagery after the storm. However, we are agreed that the figure it is not important for the story telling of the paper and we removed it.

Change of with in

R: Done changed

Formula 1

R: We changed now it is correct

why would you use such a computationally intensive process when substituting a nearby date may be easier and more similar than a linear regression predicting the SR. “The two indices were calculated because the NBR was used to detect damaged forest areas, and the NDSI to obtain information on snow cover to reduce noises”

R: We use a despike approch that is a classical approch when there are noises values in a time series. In details, we tested different way to fill the gap in correspondence of spikes. We found that snow cover have the Hight impact on NBR so we decide to use a despike when snow was present. We used NDSI to remove from the analysis the NBR values when snow cover was present in on the ground and to fill the gaps in the time series we modeled as linear regression the values of NBR in correspondence of snow cover using the two closes value (in terms of time) available.  Also in other paper other authors used different way to despike the noises in the time series see for example

To be more clear we clarified the sentence “To reduce spikes due to snow covers in the NBR TS we used a despike approch [47] in correspondence of high values of NDSI. In details, when NBR spikes were observed in correspondence of high values of NDSI (i.e., NDSI>0.1), we modeled the NBR value using a linear regression model with the two closer NBR values without snow (i.e., before and after). In fact, from Figure 3 it is possible to observe that NBR anomaly spikes (i.e., light blue line) are present when NDSI is higher than 0.1 (i.e., black points).”

Old Figure 4 new Figure 3 - what is the orange outlined box?

R: In the caption was already reported the mining of orange box. It is the time when it is possible to observe for damaged areas a deviation from seasonality. However we clarify better in the caption the mining “The orange rectangle identify the time after the storm when it is possible to observe for damaged area a deviation from the seasonality (Panel A), while undamaged area (Panel B) has no deviation from the seasonality of the two years before the storm.”

administrative Province

 

R: Thanks now it is “provinces”

suggesting a restructure to put the methods that appear before this, in the methods section.

 

R:

continues

 

R: Changed with “continuous”

above this was described as section 2.2, now 3.1. please clarify.

R: Thanks, yes it was section 2.2 Now we changed it with the right reference

“The frequency v was set to 365 for annual daily observations for the NBR time series. The iteration process is initialized by estimating the seasonal component  using the…” - but there were not 365 observations? How does this square with your every 5 days observations of Sentinel 2?

 

R: yes, you are right the number was set equal to 73 that is the number of images expected to be acquired in the area in one years.

Determinate

R: changed with “determine”

How far away from the surface relfectance values is this analysis? So far the Sr data has been despked and linear regressed, now it has been broken into three components and then those components are now being compared? This seems like a lot of processing, where the original data may have been just as well. That would be a real test - toshow the actual utility of all this processing vs not not doing it.

 

R: Yes true, the computational work appears really intensive. However, the time series we used as input data for BEAST and CCDC it is not away from the surface reflectance since we modeled just the NBR spikes in correspondence of NDSI time series, while the other values of TS were not modeled and directly derived by the images. So only in correspondence of spikes due to snow cover we used a modelled NBR value calculate as linear regression between two closer NBR values without snow covers.

Then on the basis of despiked NBR TS - BEAST detected the three components. (e.g. Wu, L.; Li, Z.; Liu, X.; Zhu, L.; Tang, Y.; Zhang, B.; Xu, B.; Liu, M.; Meng, Y.; Liu, B. Multi-Type Forest Change Detection Using BFAST and Monthly Landsat Time Series for Monitoring Spatiotemporal Dynamics of Forests in Subtropical Wetland. Remote Sens. 2020, 12, 341. https://doi.org/10.3390/rs12020341).

We also tested the possibilities to use BAEST and CCDC without despike NBR when snow cover was present. However, since the snow cover differs among different years, the noises introduced by snow do not allow to obtain accurate results since the algorithm detected as changes area where snow covers differs among different years. In our case, because forest damages where the target of the analysis, we removed the NBR spike when snow cover was present to detect only the changes due to vegetation and not due to snow. We added a sentence in the discussion to clarify it.

suggest using remainder, which is likely the word you are looking for this appears earlier in the text and later

 

R: Yes, this was a typos. Thanks a lot! We changed it in all the manuscript.

This is not sufficient. At least please give the reader some understanding of what BEAST does to the data. What does BEAST output? What does it assume? The reader does not need to know the details of the algorithm, but some background is required.

 

R: we try to clarify  adding more details “The BEAST was applied to each pixels and its own time series to detect, if present, a breakpoint of the seasonal component after the storm using the equation describe be-fore. So, applied the BEAST every month after the storm, we obtained a classification as “damaged” or “undamaged” for each pixel of the study area. At the end we obtained for each month after the storm maps of damaged and undamaged area that were used to assess the accuracy of forest windstorm detection (section 3.2) and to assess the dam-aged forest areas using the Probability-based stratified estimators (section 3.4).”

 

why 2? why not 4? If the time series is so dense, theoretically it could be as high as 100 if it was actual change?

 

R: we tested different parameters to set up the values. We used two consecutive observation to set changes since we obtained

Suggest describing this training data earlier in the paper.

 

R: we described that training data in the training dataset. However, to clarify it we added a clear reference to the section where we described the polygons “In our case the coefficients of time series models were calculated on the basis of the training sample (polygons describe in section 2.3.)”

 

require more description of RFC.

R: we added more details on RFC please see line.

 

So, if you are not submitting the BEAST changes to RFC, then what is being compared? It seems as though you are comparing two different things entirely.

R: We compared two maps derived by two different algorithms. BEAST do not required a RFC classifier while that is mandatory when we applied CCDC. For that reason, we clarify that we compared the accuracy of the maps we derived by the two approaches.

 

either you have not provided enough information to the reader, or you are comparing two different things.

R: At the end the two algorithms provide for each month after the storm a map with two classes “damaged” “undamaged” forests. So, we are comparing the accuracy of the two maps obtained.

 

suggesting confidence

 

R: change confidential with “confidence”

 

 

This item should be considered:

1) optical data respond mostly to canopy (leaves)

2) after windthrow events, leaves are largely still on the thrown logs, th sheared tops, and present in the canopy even if they are functionally dead and damaged

3) it is not at all out of the realm of possibility that in month 1 the error is is high because the leaves are still reflecting.

4) the true estimate of damage would be the next leaf on season (occurring after the event)

5) this has been fairly well proven in the literature, which is why most studies that examine forest change events tend to focus on annual best available pixel composites.

 

R: These considerations was already reported in the first version of the manuscript in the discussions. We improved it integrating with new sentences.

 

For clarity sake, you went through the intense processing to reduce linear regressed de spiked CCDC  and also the BEAST remaindered data to then reduce the categories of change to damage/no damage?

Again, this seems like overprocessing the data, when a simple solution could have worked as well (again, we have no baseline here to compare it to). For clarity sake, you mapped change/no change at 97% accuracy. This is a result yes, but why? Why could you not map severity classes? Perhaps I am misunderstanding, and for that I am sorry, but if I am it is because the text does not make this clear to the reader.

 

R: Yes we know that the a simple change detection approch may could reach the same results than the use of the algorithms we tested. However, no one have tried them for the best of our knowledge to detect forest windthrown areas. We added in the conclusion a sentence regarding that aspect. Moreover, we are agreed that severity classes are really important. However, to map severity classes we need a validation dataset. Unfortunately, we do not have these types of data. For that reason, we decided not to map just damaged and undamaged forests, but we give an estimate of damage and undamaged classes using a probability-based stratified estimator that allow to obtain an area estimate with a Standard Error and a Confidential Intervals.  Moreover, to evaluate the accuracy we also used an index that is more susceptible to the misclassification of positive cases. In fact,  we used the gmean to evaluate the accuracy of the classes that Is more appropriate than an overall accuracy. In fact, we used gmean (which is estimated from a confusion matrix as:

where TP is the number of true positives, FP is the number of false positive pixels, FN is the number of false negative pixels and TP is the number of true positive pixels) because a small gmean is an indication of poor performance in the classification of the positive cases, even if the negative cases (TN=undamaged forest) are correctly classified. So, using this index we  avoids overfitting of the negative class and under fitting the positive class.

 

 

no CCDC accuracy for month 12?

 

R: there was an error now the results was added in the graph

 

NUT?

 

R: NUT are the Nomenclature des Unités territoriales statistiques at European Level – The classification of administrative territories. We leave it in the images, but we added in the caption the description of NUT 2 and NUT3

 

Old Figure 6 new Figure 5-  NUT?

Cartographically, you should not use the same colour more than once in the map to denote different things. In this case, yellow is the damage by one algorithm color and then also the inset frame?

 

R: thanks for the suggestion we changed the color of the box. NUT we changed with Regions and Province

 

colon not needed here

 

R: we removed the colon

 

At

 

R: we changed with “To the best of”

 

Represent

 

R: we added the s now is represents

 

Hitted

 

R: changed with “hit”

 

S2

 

R: we added a clarification S2 was defined in the introduction. We added “Our results confirm”

 

the reminder components

 

R: changed with remainder

 

run on sentence. Also, appreciated is not used correctly

 

R: we revised the sentence and we change the term appreciated with detected “Firstly, in coniferous forests fallen trees remained green on the ground for a couple of months after the storm [8,14], so differences in spectral trajectories with a pixel size of 10 to 20 m cannot be immediately detected, while in broadleaves forests (i.e., mainly beech) differences in photosynthetic activities in winter between fallen trees and not damaged trees cannot be detected because tree are leafless [8]”

 

R: we rephrase the sentences

 

Sostantially

 

R: changed with substantially

 

suggest windthrown

R: we follow your suggestion and we changed “windthrowed” with windthrown

 

Conclusion - is there a reason why this novel conclusion style cannot be a paragraph? Seems like an odd choice distracting from the message.

 

R: we usually wrote our conclusion as bullet points. However, we followed your suggestion and we changed as paragraph.

 

Conflicts of Interest: “The authors declare no conflict of interest.”

 

R: we removed “ ”

 

Author Response File: Author Response.docx

Reviewer 2 Report

The amount of time and consideration that went into this research is clearly presented. I found several sections in the methods that were particularly well articulated. In general, there are only specific instances throughout the paper which could be improved to provide a better understanding for the reader. These are mostly regarding the flow of the writing, grammar, and visual accessibility of the text and figures. In the attached draft, I have highlighted the suggested grammatical revisions in Yellow and the referenced questions (below) in Blue.

 

1. Regarding the run-on sentences on pages 3 and 12. There a few sentences in both the Introduction and the Discussion which combine several thoughts and continue on for 3-4 lines or more (highlighted in Yellow). These seemed to break the flow of the material, leading to the reader having to revisit the section multiple times to understand the message.

 

2. The amount of damaged forests (42.525 ha) in the beginning of the “Materials” section (Page 3). This seems like a significantly small portion of the overall study area. How are these areas defined? Similarly on Page 5, it refers to areas “outside [of] the forests”. A definition for a minimum area for forests should be given along with a parameter such as canopy cover or basal area cover. It is not clear if these are single trees, small gaps, or 1 ha+ minimum mapping units.

 

3. The assumptions of the validation dataset. At the end of section 2.4 a statement is givens that “In the following steps of this work we considered the validation dataset as error free”. Even with a simple dichotomy classes (which this case arguably is not), these methods do not present truth, as they cannot be 100% error free (i.e., truth). While the use of high-resolution imagery photo interpretation and field sampling for the collection of reference (validation) data is a common practice that many rely on, this statement seems to present the wrong characterization of this framework.

 

4. Reference to Section 3.1 at the beginning of section 3.1.1. It seems like this is discussed more in section 2.2?

 

5. For the thematic accuracy assessment section (3.2), there should be a citation given for this practice.

 

6. Equations on Pages 7, 8, and 9. Several equations are given within the methods, they may be easier to read if there was additional spacing between the lines. This is especially true when there are multiple equations in a row.

 

7. Figures 6 (and Figure 1). The color scheme (symbology) and presentation of the map elements (e.g., North Arrow, Scale Bar, and Legend) make these Figures difficult to interpret. The map elements are almost inaccessible. Even when zooming in on the screen, the extent indicators are difficult to visualize and the text within the legend is hard to separate. Choosing a different color scheme, resolution, or sizes would make these figures much more effective.

 

Comments for author File: Comments.pdf

Author Response

Review 2

 

Thanks a lot for your comment and revision. The quality of the manuscript and Figure was improved a lot following your suggestion!

The amount of time and consideration that went into this research is clearly presented. I found several sections in the methods that were particularly well articulated. In general, there are only specific instances throughout the paper which could be improved to provide a better understanding for the reader. These are mostly regarding the flow of the writing, grammar, and visual accessibility of the text and figures. In the attached draft, I have highlighted the suggested grammatical revisions in Yellow and the referenced questions (below) in Blue.

 

1. Regarding the run-on sentences on pages 3 and 12. There a few sentences in both the Introduction and the Discussion which combine several thoughts and continue on for 3-4 lines or more (highlighted in Yellow). These seemed to break the flow of the material, leading to the reader having to revisit the section multiple times to understand the message.

 

R: Yes, we are agreed. We reorganized the introduction and discussion following your suggestion. We reported in the next replies the details of the changes we done. The changes we done following you suggestions are underlined in green.

 Reported at European Level

R: we rephrase the sentence see LINE 45-46

to support forest operations

R: we rephrase the sentence see LINE 50-51

such as Landsat (30 m) and Sentinel-2 (S2) (up to 10 m)

R:we rephrase “Landsat ( spatial resolution: 30 m, revised time: 16 days) and Sentinel-2 (S2) (spatial resolution; up to 10 m, revised time: 5 days)” LINE 71-73

the commonly

R: changed “Usually to map forest disturbances common change detection (CD) methods with optical data consist of analyzing the pre- and post-disturbance images are used” LINE 76

last years, several

R: we removed in the last years LINE 97

For this reason in the last years several applications to map forest changes were developed using multispectral optical satellite TS (i.e., Landsat, MODIS, and in the last year Sentinel-2 data) [21,27–32] thanks also to the devel-opment of cloud computing platform such as Google Earth Engine (GEE) that gives access to a complete catalog of remotely sensing data, with the capability to process efficiently large multitemporal set of remotely sensed data [33].

R: we rephrase the sentences try and we reorganized part of the introduction to be more consistent in the story telling following your suggestions. Please see LINE 97-105

To do so, we used a three years S2 Normalized Burn Ratio (NBR) time series (i.e., 01/2017-10/2019) and we evaluated the accuracy of the two algorithms using a ground true dataset calculating a probability-based stratified estimator of the total damaged area and the Standard Error of the damage area estimates, as request in the context of international reporting.

R: we rephrase all the sentences to be more precise following your suggestions. Please see line 128-138

Hitted

R: changed with hit LINE 148

2. The amount of damaged forests (42.525 ha) in the beginning of the “Materials” section (Page 3). This seems like a significantly small portion of the overall study area. How are these areas defined? Similarly, on Page 5, it refers to areas “outside [of] the forests”. A definition for a minimum area for forests should be given along with a parameter such as canopy cover or basal area cover. It is not clear if these are single trees, small gaps, or 1 ha+ minimum mapping units.

R: yes, we are agreed that this information it is important! Thanks a lot for the suggestion. The definition used in the preliminary analysis to detect damaged area was equal to the ones that we used for the validation dataset (IUTI POINT).  Minimum mapping units 2000 m² and at least 50% of trees fallen. We added this information in LINE 152-154.

Moreover, we added information on Forest definition in the introduction see LINE 128. In Study areas LINE 144 and in Sentinel2 -Time series pre-processing LINE 170.

3. The assumptions of the validation dataset. At the end of section 2.4 a statement is givens that “In the following steps of this work we considered the validation dataset as error free”. Even with a simple dichotomy classes (which this case arguably is not), these methods do not present truth, as they cannot be 100% error free (i.e., truth). While the use of high-resolution imagery photo interpretation and field sampling for the collection of reference (validation) data is a common practice that many rely on, this statement seems to present the wrong characterization of this framework.

 R: Yes, true we know that our dataset it is not error free. However, we have not the ability to calculate the error associate with the validation dataset. We removed the sentence from the file

4. Reference to Section 3.1 at the beginning of section 3.1.1. It seems like this is discussed more in section 2.2?

 

R: yes, it was an error. Now the section are renumerated according to their presentation.

 

5. For the thematic accuracy assessment section (3.2), there should be a citation given for this practice.

 R: Thanks, true. We added a citation

6. Equations on Pages 7, 8, and 9. Several equations are given within the methods, they may be easier to read if there was additional spacing between the lines. This is especially true when there are multiple equations in a row.

 R: thanks for the suggestion we revised the formula and now are spacing

7. Figures 6 (and Figure 1). The color scheme (symbology) and presentation of the map elements (e.g., North Arrow, Scale Bar, and Legend) make these Figures difficult to interpret. The map elements are almost inaccessible. Even when zooming in on the screen, the extent indicators are difficult to visualize and the text within the legend is hard to separate. Choosing a different color scheme, resolution, or sizes would make these figures much more effective.

 

R: we revised the figure in accordance also with the Review 1.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

n/a

Author Response

Dear Editor, 
thanks a lot for you suggestion. 
We did not know the paper of  Puhm et al.. Thanks a lot four your suggesiton. We incluced it in the introduction and in dissussion. Our results showed similar accuracy as the one obtained in Austria and Malawi using a different continues change detection algorithm. 
Moreover, we included also two new paper on VAIA. 
Attached in the track-change version we reported in yellow the revision you requested.