Integrating Landsat Time Series Observations and Corona Images to Characterize Forest Change Patterns in a Mining Region of Nanjing, Eastern China from 1967 to 2019

Round 1

Reviewer 1 Report

I appreciated the authors’ efforts in improving the readability of their manuscript. They clarified the many steps of the proposed methodology and provided an explanatory workflow diagram in Figure 2. However, some points need to be further improved. In the following I list the main points, and I give further suggestions on many other minor points in the annotated pdf of the manuscript.  

1) The resampling procedure for Corona data is a downsampling, as I understand. The original resolution of about 2-3 m is rescaled to 30 m. Then, what the authors call a “nearest neighbour” procedure is probably some sort of voting among the different high resolution pixels composing the new (lower resolution) one. This should be clearly stated, and the related information loss highlighted.

2) The comparison procedure (lines 250-261) is described in a very confused way. The authors should rewrite that part, to help the readers understand what is done.

3) The subsection on the Predictor Variables needs to be improved. I add several annotations on the pdf, to help the authors in their rewriting. My main points are:

- the predictive variables are not “six surface reflectance data, spectral indices, tasseled cap transformation, principal component analysis (PCA) and texture information” but instead  “texture information derived from surface reflectance data in six spectral bands, spectral indices, tasseled cap transformations, principal components of the data”, am I right?

- PCA is not simply a data reduction technology. It is an orthogonal linear transformation of multidimensional data (i.e. a change of basis) aimed at obtaining uncorrelated information on the new components. It is done by diagonalizing the covariance matrix of the original data, so that the new components are ordered according to the information content they yield. Then, PCA can be used for dimensionality reduction by projecting each data point onto only the first few principal components to obtain lower-dimensional data while preserving as much of the data variation as possible.

Comments for author File: Comments.pdf

Author Response

Dear reviewers,

 

Thank you very much for having our paper entitled " Integrating Landsat time series observations and Corona images to characterize forest change patterns in a mining region of Nanjing, eastern China from 1967 to 2019" reviewed and sending us a bunch of comments, which are quite helpful for us to improve the manuscript. We have carefully revised the manuscript in accordance with the comments raised in the peer-review process, and all changes have been highlighted in the manuscript. Additionally, our corresponding point-by-point replies are presented in the following. Lastly, we also have reorganized and rewrote relevant sections, improved English and checked the entire sections of the manuscript including main text, figures, tables and references to ensure its compliance with the style or format of Remote Sensing.

 

We hope this version is acceptable and we are grateful for your reconsideration and looking forward to your positive responses.

 

Response to Reviewer 1 Comments

I appreciated the authors’s efforts in improving the readability of their manuscript. They clarified the many steps of the proposed methodology and provided an explanatory workflow diagram in Figure 2. However, some points need to be further improved. In the following I list the main points, and I give further suggestions on many other minor points in the annotated pdf of the manuscript.

Response: Thanks for your points, especially the suggestions you mentioned in the pdf, which helped us a lot improve the manuscript. Based on your comments, we have made some modifications accordingly. All modifications are marked in red ink in the revised version. The itemized response to each comment is provided as follows:

1) The resampling procedure for Corona data is a downsampling, as I understand. The original resolution of about 2-3 m is rescaled to 30 m. Then, what the authors call a “nearest neighbour” procedure is probably some sort of voting among the different high resolution pixels composing the new (lower resolution) one. This should be clearly stated, and the related information loss highlighted.

Response: Thanks for your suggestions. We strongly agree with your comment. Downsampling the 2-3m forest cover maps of orthorectified Corona images to 30-meter maps, was done again by implementing a “majority filter” instead of “nearest neighbour”. Thus, the code of the lower resolution pixels can be reasonably determined. The corresponding results has been modified. The text has been revised as “Finally, the forest cover images were downsampled or aggregated to 30 m resolution by considering the area-dominated criterion (majority rule) [22]. Some narrow road information may be lost during downsampling” (section 2.3.2, line 225 to 227)

2) The comparison procedure (lines 250-261) is described in a very confused way. The authors should rewrite that part, to help the readers understand what is done Response: Thanks for your suggestions. We have rewritten the part as: “Identifying forest loss/recovery process categories were based on the number of unchanged forest patches and non-forest patches around the lost/gained forest patches. Firstly, lost forest pixels, unchanged forest pixels and non-forest pixels were obtained by comparing the bi-temporal forest/non-forest products, where the unchanged forest was set to 1, the non-forest was set to 0, and the lost forest was set to null. Then, we reassigned different values to each unchanged forest patch (starting with the number 1). Next, we calculated the maximum value of each pixel in the lost patches within the 8-pixel neighbourhood and assigned this maximum value to the central pixel. Finally, we counted the number of different values in each lost patch. If the lost forest patch contained two different values, then the patch was shrinkage. If the lost forest patch contained three or more different values, then the patch was subdivision. If the lost forest patch contained only one value, and the value was 0, then it experienced attrition, while the remaining patches were perforation patches. The forest restoration process model was similar to the forest loss process model, except that the loss patches were replaced with gained patches.” (section 2.3.4, line 260 to 272)

3) The subsection on the Predictor Variables needs to be improved. My main points are:the predictive variables are not “six surface reflectance data, spectral indices, tasseled cap transformation, principal component analysis (PCA) and texture information” but instead “texture information derived from surface reflectance data in six spectral bands, spectral indices, tasseled cap transformations, principal components of the data”, am I right? - PCA is not simply a data reduction technology. It is an orthogonal linear transformation of multidimensional data (i.e. a change of basis) aimed at obtaining uncorrelated information on the new components. It is done by diagonalizing the covariance matrix of the original data, so that the new components are ordered according to the information content they yield. Then, PCA can be used for dimensionality reduction by projecting each data point onto only the first few principal components to obtain lower-dimensional data while preserving as much of the data variation as possible.

Response: We are sorry for making the description unclear. (1) The potential predictive variables included both six surface reflectance bands and other derivatives including the spectral indices, tasseled cap transformation, principal component analysis (PCA) and texture information derived from the original surface reflectance bands) (section 2.3.5, line 276 to 279).

(2) We are very grateful for your explanation of PCA, we have rewritten the definition of PCA as “PCA is an orthogonal linear transformation of multi-dimensional data, which is completed by diagonalizing the covariance matrix of the original data. Then, PCA can be used for dimensionality reduction by projecting each data point onto only the first few principal components to eliminate redundant information without losing important information and it attains a possible closer relationship with AGB than a single spectral band” (section 2.3.5, line 285 to 290).

4) I add several annotations on the pdf, to help the authors in their rewriting

Response: Thanks for your comments in the provided pdf, which greatly helped us improve the quality of the manuscript. In response to our unclear sentences, incomplete figure caption and grammatical problems, we have made corresponding changes. Some of your doubts are explained below:

(1) Regarding your question about "first" in Figure 7, we make the following response: a pixel may experience multiple disturbances or recovery during the entire time series analysis, we only recorded the year and location of the first disturbance or recovery (section 3.2, line 355 to 356). Besides, you mentioned the question about the blank area in Figure 7, we take the Fig. 7 (a) as an example to explain it. The disturbance map only recorded unchanged categories and disturbance categories. Therefore, the blank area represented the areas where forest restoration occurred. Similarly, the blank area on the restoration map referred to the location of forest disturbance. We have replotted Figure 7 and added a legend about the blank area.

(2) Regarding your question about the accuracy evaluation of forest cover products in the pdf, we mentioned it in section 2.3.3 “field survey data and forest cover maps in 2006 and 2017 were exploited to obtain the producer, user and overall accuracies to validate the forest cover products. Producer accuracy refers to the ratio of the correctly classified samples to the samples on the real map for a certain category. User accuracy refers to the ratio of the samples correctly classified to the samples on the classified image for a certain category. Overall accuracy refers to the ratio of the samples correctly classified to the total samples for all categories [42].” (section 2.3.3 line 238 to 244).

Besides, We have revised the relevant result description according to your suggestion:

“By comparing the forest cover maps for 2006 and 2017 with the inventory data, we found that the overall accuracies of the two-year classification results were above 90%, and the highest overall accuracy was 94.8% in 2006, while the lowest overall accuracy was 93.0% in 2017” (section 3.1 line 334 to 338)

 

Reviewer 2 Report

The authors used remote sensing-based AGB and CORONA AGB to analyze the ecosystem change between 1970 and 2017. I still have one concern about the revised paper, this is although the authors have shown Landsat derived AGB has good accuracy, that product still has high uncertainty, so that the author must address this uncertainty for the AGB comparison. In addition, the authors cited two other papers (reference 28 and 29) about using field survey data and remote sensing data for biomass comparison, I think the use of these two data of different data source for comparison is because (1) there is no remote sensing data available at 1970s, or (2) field plot data is not available after 2000. so that the use of different data sources is just a optional choice, although not perfect. However, the author just claims "the comparison of the two data sources is reasonable (line 491~493, and line 514~516) because other people also do this", this kind of statement is not acceptable.   

Author Response

Dear reviewers,

 

Thank you very much for having our paper entitled " Integrating Landsat time series observations and Corona images to characterize forest change patterns in a mining region of Nanjing, eastern China from 1967 to 2019" reviewed and sending us a bunch of comments, which are quite helpful for us to improve the manuscript. We have carefully revised the manuscript in accordance with the comments raised in the peer-review process, and all changes have been highlighted in the manuscript. Additionally, our corresponding point-by-point replies are presented in the following. Lastly, we also have reorganized and rewrote relevant sections, improved English and checked the entire sections of the manuscript including main text, figures, tables and references to ensure its compliance with the style or format of Remote Sensing.

 

We hope this version is acceptable and we are grateful for your reconsideration and looking forward to your positive responses.

 

Response to Reviewer 2 Comments

 

The authors used remote sensing-based AGB and CORONA AGB to analyze the ecosystem change between 1970 and 2017. I still have one concern about the revised paper, this is although the authors have shown Landsat derived AGB has good accuracy, that product still has high uncertainty, so that the author must address this uncertainty for the AGB comparison. In addition, the authors cited two other papers (reference 28 and 29) about using field survey data and remote sensing data for biomass comparison, I think the use of these two data of different data source for comparison is because (1) there is no remote sensing data available at 1970s, or (2) field plot data is not available after 2000. so that the use of different data sources is just a optional choice, although not perfect. However, the author just claims "the comparison of the two data sources is reasonable (line 491~493, and line 514~516) because other people also do this", this kind of statement is not acceptable. 

Response: We are sorry that we did not clearly introduce the related data in AGB part. We have modified the relevant description to make readers better understand the remote sensing variables involved in the AGB model. In this study, we only predicted AGB in 2006 and 2017 based on Landsat images and field survey data. Corona data was not involved in the AGB modelling and they were only used to reconstruct forest cover histories and landscape structure in 1967 and 1970.

As for the uncertainties of the AGB prediction models, they are mainly attributable to the Landsat data source. We all know that Landsat provides relatively rich spectral signals on forest canopy to describe the horizontal structure of forest canopy, which can not penetrate the canopy to provide the vertical structure information. Thus, sole use of Landsat to predict forest AGB usually bring about the spectral saturation effect, leading to an overestimation of the low AGB observations while an underestimation of the high AGB observations. To address the effect, accurate description of the vertical structure of forests by using microwave even LIDAR signals, or certain ground-based forest structural attributes, is needed in the modelling process, for example, the crown density used in the current modelling work. These can be done if these data are available or enough time and money are used to collect this data. In our current work, we only used the free Landsat data coupled with field sample data to map forest AGB, thus, the existence of uncertainties can be easily understood. Currently, it has become a popular method to predict AGB by coupling Landsat and field survey data and there are also references proving the effectiveness of this technique [27-29] (line 545 to 547). Besides, the new ideas about AGB part in this study is to combine the landscape structure and ecological functions to reflect the performance of the mining area restoration project and concluded that forest restoration has increased structural connectivity, but its ecological function is still low (Table 9) (section 3.4, line 432 to 433).

In the previous manuscript from line 491 to 493, which are line 523 to 524 (section 4.2) for the current version, we may not clearly express it. This sentence was used to express that the practicability of the forest spatial process model is not limited to our research area, it can be applied to any land cover change map in other study areas.

 

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript was improved after the revision. However, the "Results" section needs more work. I suggest authors re-write the Results section of the manuscript taking into consideration these specific suggestions on "Methods" section:

Line 121 – there are no data about the topography, slope distribution, as it is well known that high energy topography is influencing the orthorectification process

Line 151. – no accuracy assessment was done on the corona image. What was the average error, distribution of errors. Please add the assessment in an annex (errors per GCP on x,y,z or errors on random points)

Lines 169 – 170: Which sampling methodology was done in the field visit? It is well known in field inventory that based on the variation coefficient and the target accuracy a number of samples is calculated (especially for stocking volume, average DBH and stand height)

Line 172 – more information about the polygons is needed. There are serious doubts that the estimations can be biased (spatiality and homogeneity). Please add a map with their distributions in the annex

Line 283 – it is not clear which of the field data were used to train the random forest and make the regression – additional data need to be added as an annex

380 – how the 21  samples for 2006 and 32 samples for 2017 were selected?

General comments:

Abbreviations should be explained at the first use

Author Response

Dear reviewers,

 

Thank you very much for having our paper entitled " Integrating Landsat time series observations and Corona images to characterize forest change patterns in a mining region of Nanjing, eastern China from 1967 to 2019" reviewed and sending us a bunch of comments, which are quite helpful for us to improve the manuscript. We have carefully revised the manuscript in accordance with the comments raised in the peer-review process, and all changes have been highlighted in the manuscript. Additionally, our corresponding point-by-point replies are presented in the following. Lastly, we also have reorganized and rewrote relevant sections, improved English and checked the entire sections of the manuscript including main text, figures, tables and references to ensure its compliance with the style or format of Remote Sensing.

 

We hope this version is acceptable and we are grateful for your reconsideration and looking forward to your positive responses.

 

Response to Reviewer 3 Comments

 

The manuscript was improved after the revision. However, the "Results" section needs more work. I suggest authors re-write the Results section of the manuscript taking into consideration these specific suggestions on "Methods" section: 

Response: Thanks for your suggestion, we have added the corresponding table and figure according to your suggestion to make the manuscript more readable. All modifications according to reviewers’ comments are marked in red ink in the revised version. The itemized response to each comment is provided as follows:

(1) Line 121 – there are no data about the topography, slope distribution, as it is well known that high energy topography is influencing the orthorectification process 

Response: Thanks very much for pointing this out. When performing ortho-rectification in previous manuscripts, we have already considered DEM data, but we are sorry that we did not describe it in the Table 1. We have added a description of topography data “DEM data were from the NASA Shuttle Radar Topographic Mission (SRTM) to implement subsequent ortho-rectification.” (section, 2.2.1, line 138 to 139) (Table 1)

(2) Line 151. – no accuracy assessment was done on the corona image. What was the average error, distribution of errors. Please add the assessment in an annex (errors per GCP on x,y,z or errors on random points)

Response: Thanks for your suggestions, we have inserted a table and corresponding text to evaluate the accuracy of ortho-rectification. See the manuscript for the table 5, and the text is as follows: “We generated 11 GCPs on the orthorectified Corona image and Landsat to assess the errors of ortho-rectification. The errors of all GCPs were controlled within half a pixel (15 m) (Table 5). So the geographic accuracy of Corona data was comparable to Landsat (Figure 5), and there were few false changes caused by errors in ortho-rectification (Table 5) ” (section 3.1, line 330 to 333)

(3) Lines 169 – 170: Which sampling methodology was done in the field visit? It is well known in field inventory that based on the variation coefficient and the target accuracy a number of samples is calculated (especially for stocking volume, average DBH and stand height) 

Response: We are sorry for making it unclear in the field data section. In this research, we used forest management planning inventory (FMPI) and it is committed to assessing forest resources and to supply information requirements for forest management planning, spatial and functional patterns, which are completely consistent with our research goals. The primary methods of the FMPI include photography interpretation (subcompartment division), simple visual estimations, and angle-gauge sampling. The estimation of subcompartment inventory is compared from the estimation of overall sampling survey based on systematic sampling. If the difference does not exceed ±1 standard error, the subcompartment stocking volume will be regarded as correct. If the difference is between standard error and three times standard error, error sources will be investigated, analyzed, and corrected until the difference is within the single standard error range. In the case where the difference is greater than three times the standard error, the subcompartment stocking volume must be measured again. Thus, we think that FMPI is reliable in this study.

The text is revised as follows: “Mt. Mufu’s forest management planning inventory (FMPI) [36, 37] - the second level of China’s forest inventory system, was conducted in 2006 and 2017. This type of data divides the forest areas into multiple subcompartments and is stored in vector form. Based on simple visual estimations and angle-gauge sampling, the investigator recorded a bunch of forest characteristics in each subcompartment, including the dominant tree species, origin, crown density, age, site quality, average forest height, average diameter at breast height, stocking volume per unit area, slope and aspect etc.” (section 2.2.2, line 171 to 177)

Once these attributes were derived from each subcompartment, we randomly picked up 80%-85% field data for AGB model training and the remaining field data for model validation.

(4) Line 172 – more information about the polygons is needed. There are serious doubts that the estimations can be biased (spatiality and homogeneity). Please add a map with their distributions in the annex 

Response: Thanks for your suggestions. We have added a plot distribution map in the study area (Figure 1 (d)) and hope it can help solve your doubts. It can be seen from figure 1 (d) that the distribution of plots is relatively uniform. And the calculation of AGB came from each subcompartment. A subcompartment is a contiguous forest region that is quite homogeneous and contains a bunch of forest features. Therefore, we think that the evaluation errors due to spatiality and homogeneity will be relatively small

(5) Line 283 – it is not clear which of the field data were used to train the random forest and make the regression – additional data need to be added as an annex 

Response: We are sorry for not explaining this part clearly. The input variables of the field data used for the AGB calculation were stocking volume per unit area and dominant tree species. “We first adopted the biomass conversion factor method [37, 38] to convert the stocking volume into the forest AGB according to different dominant species in each subcompartment (Appendix Table A1)” (section 2.2.2, line 179 to 181)

The AGB formula is as follows: AGB=a+b*V

where a and b are constants for a certain forest type, and V is the stocking volume per unit area. We reinserted Appendix Table A1.

Besides, to incorporate the crown density and other remote sensing variables as predictors to evaluate AGB, the polygon crown density was converted into raster images as the predicted variable (line 182 to 184).

(6) line380 – how the 21 samples for 2006 and 32 samples for 2017 were selected? 

Response: We are sorry for not explaining this part clearly. In section 2.2.2, we summarized that there were 104 subcompartments in 2006 and 187 subcompartments in 2017. 80%-85% of the field data were randomly selected for model training, and the remaining data is used for model validation. (line 184 to 185)

General comments:

Abbreviations should be explained at the first use

Response: Thanks for your suggestions. We have made a brief explanation of the main abbreviation.

(1) rational polynomial coefficients (RPC): RPC is used to establish the geometric relationship between the plane coordinates of the pixel and the corresponding ground point spatial coordinates. RPC was generated by setting camera parameters and selecting ground control points (GCPs) with DEM data to establish external orientation parameters (line 159 to 162)

(2) vegetation change tracker (VCT): “VCT is an automated analysis algorithm based on the spectral-temporal characteristics of the forest change process in the time series to reconstruct forest disturbance and recovery history” (line 190 to 193)

(3) The interpretation of remote sensing variables related to AGB modeling was described in section 2.3.5 (line 282 to 295)

(4) ntree, mtry ,nodesize: The ntree is the number of decision trees and 500 trees (ntree) were used for AGB modelling. The mtry is the number of variables to be tested at each node and it was set to 4. The nodesize parameter is the minimum number of decision tree nodes and it was set to the default value of 1 in the current work. (line 302 to 305)

(5) PercentIncMSE is defined as the decrease in accuracy when a given variable was excluded from decision trees. IncNodePurity measures the decrease in node impurities attributable to the splits of a given variable [29]” (line 309 to 311)

 

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

All of my comments have been addressed and related modifications to the manuscript have been done.

I suggest to accept the manuscript for publication, provided that these minor points are modified:

Line 181: 80% - 85% of the field data were randomly selected for model training

I don't understand: 80% or 85%? If you are referring to two different choices for the different datasets you should specify the reason for this choice.

Line 258: Identifying forest loss/recovery process categories IS / WAS (not "were") based

Line 276: "derived variables", not "derivatives"

Line 276: "principal components", not "principal component analysis"

Line 341: "were obtained", not  "were obtain"

Author Response

Dear reviewers,

 

Thank you very much for having our paper entitled " Integrating Landsat time series observations and Corona images to characterize forest change patterns in a mining region of Nanjing, eastern China from 1967 to 2019" reviewed and sending us useful comments. We have carefully revised the manuscript in accordance with the comments raised in the peer-review process, and all changes have been highlighted in the manuscript. Additionally, our corresponding point-by-point replies are presented in the following.

 

We hope this version is acceptable and we are grateful for your reconsideration and looking forward to your positive responses.

 

Thank you and best Regards.

Your sincerely.

Yali Zhang, Wenjuan Shen, Mingshi Li, Yingying Lv

 

Response to Reviewer 1 Comments

All of my comments have been addressed and related modifications to the manuscript have been done.

I suggest to accept the manuscript for publication, provided that these minor points are modified: 

Response: Thanks for positive feedback. We further have revised those inappropriate descriptions and grammatical problems to make the manuscript more readable. All modifications according to reviewers’ comments are marked in red ink in the revised version. The itemized response to each comment is provided as follows:

1.Line 181: 80% - 85% of the field data were randomly selected for model training

I don't understand: 80% or 85%? If you are referring to two different choices for the different datasets you should specify the reason for this choice.

Response: Thanks for your point, we are sorry for not expressing it clearly here. We adopted the sample function of the R software to perform sampling, and set the probability of each element to be sampled to 0.8 and 0.2, so the training and validation samples did not completely obey the ratio of 0.8 to 0.2 (There is a subtle difference). Thus, we used an inappropriate description like 80%-85%. We have revised the corresponding description as “We adopted the sample function of the R software to perform sampling, and set the sampling probability of each element to 0.8 and 0.2 for model training and validation.” (line 184 to 185). The number of validation samples was shown in Figure 9

2. Line 258: Identifying forest loss/recovery process categories is/ was(not "were") based

Response: Thanks for your point, we have revised “were” as “was” (line 260)

3. Line 276: "derived variables", not "derivatives"

Response: Thanks for your point, we have revised “derivatives” as “derived variables” (line 277 to 278)

4. Line 276: "principal components", not "principal component analysis"

Response: Thanks for your point, we have revised “principal component analysis” as “principal components” (line 278)

5. Line 341: "were obtained", not "were obtain"

Response: Thanks for your point, we have revised “obtain” as “obtained” (line 343)

 

Reviewer 3 Report

I consider that all the suggestions were addressed in the present form of the manuscript

Author Response

Dear reviewers,

 

Thank you very much for having our paper entitled " Integrating Landsat time series observations and Corona images to characterize forest change patterns in a mining region of Nanjing, eastern China from 1967 to 2019" reviewed and giving us your positive responses.

 

Best Regards.

Your sincerely.

Yali Zhang, Wenjuan Shen, Mingshi Li, Yingying Lv

 

Response to Reviewer 3 Comments

I consider that all the suggestions were addressed in the present form of the manuscript

Response: Thanks for your positive comment.

 

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The authors intend to investigate spatio-temporal dynamics of forest cover in a very challenged mining area by integrating Landsat and Corona data. The subject could be interesting, and the rich literature on the assessment of forest changes has recently shown an increasing consideration for such a combined approach.  However, the proposed investigation procedure is not clearly described, and many points are not sufficiently explained. The manuscript in its present state is not mature for publication yet.

In the following I list the main improvements I suggest:

1) First (and essential) The complete processing pipeline should be schematized: in the present text it is unclear what data are used to what purpose, what indicators are evaluated, … I suggest to provide a flowchart of the whole procedure, along with a few introducing sentences for each subsection to clearly state what it is aimed at.

2) The entire procedure applied to Corona images (registration, annotation, resampling) should be described more precisely. Moreover, as I understand, the authors are not actually combining Corona and Landsat data. They use the former data  to obtain the forest initial condition,  and the latter ones to detect changes. It should be more clearly stated.   

3) All the adopted variables, and the accuracy indicators should be defined.

4) The variable selection procedure is also unclear. It seems that the authors simply run a standard R package, without reporting quantitative results. Moreover, ref [45] is wrong, I assume (or, at least, it is not about Random Forests and Variable Selection).

5) Model Accuracy Assessment and Validation. The authors discuss values of R^2 and RMSE, but they didn’t report any value.

6) Results are simply qualitatively described and commented. Also the use of statistical indicators is unclear

7) The references should be checked. Many of the authors’ names are misprinted.

Reviewer 2 Report

The paper is generally interesting as combining Landsat and CORONA data for forest cover dynamic monitoring. And also used remote sensing data to describe the forest loss and recovery process. While I just have some minor concerns:

1. I suggest the authors show some CORONA data and their visually interpreted forest distributions.

2. Authors used Random Forest to predict AGB between 2006 and 2017, and I wonder what are the training samples used for the Random Forest.

3. The study region of this study is very small, so that please comment on whether the methodology of this study applicable in other study regions of forest dynamics and ecosystem function analysis.

Reviewer 3 Report

The article has serious flaws regarding Corona ortho-rectification. The entire research must be done based on a quality orthorectified product according to international practice on this subject

• Altmaier, Angela, and Christoph Kany. "Digital surface model generation from CORONA satellite images." ISPRS Journal of photogrammetry and remote sensing 56.4 (2002): 221-235.
• Casana, Jesse, and Jackson Cothren. "Stereo analysis, DEM extraction and orthorectification of CORONA satellite imagery: archaeological applications from the Near East." Antiquity 82.317 (2008): 732-749.
• Galiatsatos, N., 2009. The shift from film to digital product: focus on CORONA imagery.Photogrammetrie-Fernerkundung-Geoinformation 2009, 251–260.
• Galiatsatos, N., Donoghue, D.N.M., Philip, G., 2007. High resolution elevation data derived from stereoscopic CORONA imagery with minimal ground control. Photogramm. Eng. Remote Sens. 73, 1093–1106.
• Gheyle, W., Bourgeois, J., Goossens, R., Jacobsen, K., 2011. Scan problems in digital CORONA satellite images from USGS archives. Photogramm. Eng. Remote Sens. 77, 1257–1264. https://doi.org/10.14358/PERS.77.12.1257.
• Kim, K., K. C. Jezek, and H. Liu. "Orthorectified image mosaic of Antarctica from 1963 Argon satellite photography: image processing and glaciological applications." International Journal of Remote Sensing 28.23 (2007): 5357-5373.
• Nita, M.D., Munteanu, C., Gutman, G., Abrudan, I.V., Radeloff, V.C., 2018. Widespread forest cutting in the aftermath of world war II captured by broad-scale historical Corona spy satellite photography. Remote Sens. Environ. 204, 322–332. https://doi.org/10.1016/j.rse.2017.10.021.
• Pollefeys, Marc, Reinhard Koch, and Luc Van Gool. "Self-calibration and metric reconstruction inspite of varying and unknown intrinsic camera parameters." International Journal of Computer Vision 32.1 (1999): 7-25.