How BFAST Trend and Seasonal Model Components Affect Disturbance Detection in Tropical Dry Forest and Temperate Forest
Round 1
Reviewer 1 Report
The disturbance detection of forest is important for the ecological protection, and during this analysis the remove of seasonal cycles is essential. This study aimed to investigate the effect of trend and seasonal components on disturbance detection in tropical dry forest and temperate forest, which is critical in the study area. However, there are still some problems as follows.
Figure 1: The monthly time series is from January 2013 to January 2016, please add the data between February 2016 and 2018.
L172-174: 1. The percentage of cloudless observations are higher in the dry season, and lower in the rainy season. This is inconsistent with many other studies, and what is the reason?
L25: Why define 1994-2015 as the history period and 2016 January - 2018 June as the monitoring period? In other words, if define “1994-2010” and “2011-2018” as the history and monitoring period, respectively; the result may be different.
L270: How determine the “stable period” in this study?
L279-280: There are no ɑ5, ɑ6, x5 and x6 in equation 5.
L292-294: Forest disturbances were about 329 ha in TDF, and 168 ha in TF; therefore, the sum of them was about 497 ha, but not 504.7 ha.
Figure 4: In my opinion, B) and D) are part of A) and C), respectively. Therefore, the author should add the arrow and description to make them clearer.
Figure 5: Some colors in the Legend are similar, making it is difficult to distinguish them.
L333-344: What is the difference between “user’s accuracy” and “producer’s accuracy”?
Figure 6: The significant correlations or not significant at which significance level?
L480-485: More detailed and valuable information should be added to the conclusion section.
Author Response
Response to reviewer 1
Thank you very much for your comments and suggestions. We have provided in the following the responses to your comments one by one inside “[]”.
Comments and responses:
The disturbance detection of forest is important for the ecological protection, and during this analysis the remove of seasonal cycles is essential. This study aimed to investigate the effect of trend and seasonal components on disturbance detection in tropical dry forest and temperate forest, which is critical in the study area. However, there are still some problems as follows.
[Thank you very much.]
Figure 1: The monthly time series is from January 2013 to January 2016, please add the data between February 2016 and 2018.
[We have added the data between February 2016 and 2018, and now the monthly time series NDVI profile for both temperate forest and tropical dry forest covers a time span from December 2012 to March 2019.]
L172-174: 1. The percentage of cloudless observations are higher in the dry season, and lower in the rainy season. This is inconsistent with many other studies, and what is the reason?
[Sorry for not being clear to cause the confusion. In the previous version we were actually showing in table 1 the percentage of cloudless observations. We have now recalculated the percentage of no data (NAs) observations for every month and updated the table 1, in which the percentage of NAs in dry season is 24% and in rainy season is 47%, which is consistent with other studies.]
L25: Why define 1994-2015 as the history period and 2016 January - 2018 June as the monitoring period? In other words, if define “1994-2010” and “2011-2018” as the history and monitoring period, respectively; the result may be different.
[We defined the historical period and monitoring period based on the availability of Landsat data and the consideration of ground verification for the detected disturbances. The historical period starts at the beginning of the dataset which is January of 1994 and ends at the end of 2015, and we defined the monitoring period from 2016 to 2018 June because we expect that the changes within 2 and a half years of time would still be somehow recognizable in the field survey, which was carried out in October of 2018. We have added the related information in the 2nd paragraphs of the methods section 2.3. Yes, the result will most probably be different if the monitoring period is longer, for example 2011-2018, since potentially more changes will be detected.]
L270: How determine the “stable period” in this study?
[The BFAST algorithm determines a stable historical period using a reverse ordered CUCUM test (ROC) (Verbesselt et al 2010). We have added this information in the 1st paragraph of the methods section 2.3.]
L279-280: There are no ɑ5, ɑ6, x5 and x6 in equation 5.
[Thanks. We have removed ɑ5, ɑ6, x5 and x6 from the equation 5 and updated the equation with the information of the four independent variables.]
L292-294: Forest disturbances were about 329 ha in TDF, and 168 ha in TF; therefore, the sum of them was about 497 ha, but not 504.7 ha.
[Thank you. In a previous version, we have used a slightly different mask to filter out the pixels that were not forests at the beginning of the monitoring period, which resulted in a slightly different number of disturbances. We have now made corrections in the revised version.]
Figure 4: In my opinion, B) and D) are part of A) and C), respectively. Therefore, the author should add the arrow and description to make them clearer.
[Thanks for the suggestion, we have revised the figure and made adjustments in both the figure and its description to make them clearer.]
Figure 5: Some colors in the Legend are similar, making it is difficult to distinguish them.
[We have revised the figure 5 using a different color pallet to make the colors more distinguishable. However, to center the article more around the objective which is the evaluation of the contribution of BFAST components to disturbance detection, we have decided to remove this figure. We hope you agree.]
L333-344: What is the difference between “user’s accuracy” and “producer’s accuracy”?
[We revised the error matrix to make the user’s and producer’s accuracy more evident for tropical dry forest and temperate forest. In the revised version, it can be clearly seen that the disturbance detection had higher accuracy in temperate forest than in tropical dry forest.]
Figure 6: The significant correlations or not significant at which significance level?
[The significance of the correlations is at 95% confidence level. We have added this information in the text, section 3.4 and in the caption of the Figure 6.]
L480-485: More detailed and valuable information should be added to the conclusion section.
[Thanks for the suggestion. We have now added more detailed and valuable information to the conclusion section.]
Reviewer 2 Report
All my remarks and suggestions can be found in main document
Comments for author File: 
Comments.pdf
Author Response
Response to reviewer 2:
Thank you very much for your comments and suggestions. We have provided in the following our responses to your comments one by one inside “[]”.
Comments and responses:
All these services are not only benefiting to the poor.
[We removed "for the world's poor" from the sentence, so that it is clear all humanity benefits from forest services.]
Consider rephrasing this sentence.
[We have re-phased this sentence as "Reliable information on forest cover and its changes is crucial for policy makers to design effective plans in forest conservation."]
This idea has already been highlighted in L46-52.
[We deleted this sentence and accommodated the references].
State the Latin name of the species.
[We added the scientific name of avocado: "Persea americana".]
Inter calibration of the different landsat sensors as proposed by Roy et al; 2016 https://doi.org/10.1016/j.rse.2015.12.024 is recommended to reduce the sensor effects on the radiometry of the scenes.
[Thanks. We revised the text by stating the importance of inter-calibration of the different sensors in a time series analysis. We also explained that in our case, we consider using un-calibrated scenes would have minor effects on the change detection since the selected historical period (1994 – 2015) includes at least 2 complete years where Landsat-7 ETM and Landsat-8 OLI were operational. Therefore, we assume that the BFAST algorithm can fit a stable model considering the NDVI variations between the two different sensors.]
What do the different colours represent? temperate forest are..... tropical dry forest ......
[We added the missing information that the temporal profile of NDVI of the temperate forest is represented by red points and that of the tropical dry forest by blue points.]
Latin name.
[We replace "agave" with its Latin name "Agave spp."]
Existing forests before the beginning of the monitoring period should instead be considered.
[Thanks, and sorry for the confusion. We have revised the text to make it more explicit that we updated a land cover map in 2011 using visual interpretation to obtain the land cover map of 2016 to make sure the existing forests before the beginning of the monitoring period are being analyzed.]
To account on the effect the magnitude of change, goodness of fit, amplitude and length of stable period, the analysis must be conducted on pixels with the same length of data availability i.e pixels with the same number of observations. This is more important to make sure that obtained differences are not related to the uneven distribution within the sampled pixels.
[Thanks for this important observation. We have revised the article by stating the importance of using pixels of even distributions in the analysis, while explaining that in our case, since the pixels are from the same study area and the time series data include a long historical period, we assume that the effects from the difference in the number of observations could be ignored. ]
The corresponding label on the figure is missing.
[We revised the text and the figure caption to make sure they correspond to the label on the figure.]
Rephrase the sentence.
[We rephrased the sentence as " The dashed vertical black line separates the observations of the historical period, represented by green points, from the ones of the monitoring period, represented by red points."]
The focus of the study is to evaluate the contribution of the BFAST model components and the percentages of pixels with NA on disturbance detection. However, from the results and discussion it seems like the authors laid more emphasis on testing the potentiality of BFAST models in predicting disturbances. Although this last objective is not the least it confuses the author and dilute the key message of the paper. First of all, no quantitative contribution of the different Bfast model components in estimating disturbances is given. the fact that some components are correlated amongst themselves does not imply that they have an influence on the model performances. To make sure that the differences observed are not related to the uneven distribution of the number of data available between pixels within a time series it is important to compare pixels that have the same length of data availability. The objectives or the results should be re-adjusted for them to match.
[Thank you for this important observation. To address this observation thoroughly, we have revised the sections of results, discussions, and conclusions. 1) in the results section, to center around the objective of the article, which is to evaluate the contributions of BFAST components to disturbance detection, we have deleted the figure 5 which shows the BFAST disturbance detection, and added a new figure 5, which shows the statistical description of BFAST components. We believe this new figure lays foundation to later analysis of BFAST components with the logistic regression. 2) in the results of logistic regression, we added a new coefficients "odds ratio" to the results of logistic regression. We believe that it helps to explain the contribution of the model components to disturbance detection. 3) in the discussion section, we deleted the section 4.1 as it centers on discussing how well BFAST model detects disturbances; we added a new section on BFAST model components, and logistic regression with all forests and with specific type of forest. 4) we expanded the conclusion section, making it clear the objective methods and conclusions are in alignment.]
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
This version has been improved significantly. However, there are some new critical problems about the data.
(1) Figure 3 (monthly time series NDVI) in this version is different obviously from that in the previous version; but, most other figures and tables in these two versions are the same. If the author have reprocessed the data, there may be some differences between most figures and tables. If not, why these is remarkable difference of Figure 3?
(2) The correlations between NA percentage and magnitude in Figure 6 are 0.457, 0.362, and 0.437 for all forests, TDF only, and TF only (Figure 6A, 6B and 6C), respectively in this version; however, they are 0.457, 0.362, 0.311 in the previous version. Why the correlation for TF has changed, but it was the same for all forests? Moreover, many correlations have the same modulus but opposite sign in Figure 6 between this version and the previous version.
Therefore, the data should be checked carefully throughout this version.
Author Response
Thank you very much for your valuable comments and suggestions. We have provided in the following the responses to your comments one by one inside “[]”.
(1) Figure 3 (monthly time series NDVI) in this version is different obviously from that in the previous version; but, most other figures and tables in these two versions are the same. If the author has reprocessed the data, there may be some differences between most figures and tables. If not, why these is remarkable difference of Figure 3?
[No data was reprocessed. In this version, we chose two different pixels to demonstrate the NDVI time series profile for temperate forest (TF) and tropical dry forest (TDF). That is why Figure 3 in this version looks different from that in the previous version.
In the previous version, we demonstrated the NDVI time series profile of TDF and TF from January 2013 to January 2016. Since the reviewer advised us to add the data in the period between February 2016 and 2018, we decided to use a different pixel to demonstrate the data profile. This is because in the previous version we used pixels where BFAST detected disturbances in the monitoring period (2016 – 2018), and if we only extend the time period, the change detected by BFAST will blur the difference in the NDVI profiles between TDF and TF. By using different pixels where no change detected we can emphasize better the differences in the time series data of these two types of forest.
This data was not included in the analysis and the only purpose of Figure 3 is to demonstrate the differences in NDVI profiles between TF and TDF.]
(2) The correlations between NA percentage and magnitude in Figure 6 are 0.457, 0.362, and 0.437 for all forests, TDF only, and TF only (Figure 6A, 6B and 6C), respectively in this version; however, they are 0.457, 0.362, 0.311 in the previous version. Why the correlation for TF has changed, but it was the same for all forests? Moreover, many correlations have the same modulus but opposite sign in Figure 6 between this version and the previous version.
[We thank the reviewer for pointing this out. The change of sign but with the same modulus results from homogenizing magnitude as the absolute value, which we described in the methods section, in 2.3, paragraph 4, line 225-227. However, the changes in the correlation for TF were caused by an incorrect subset of the data. We have updated the data and carefully checked the text to make sure they match.]
(3) Therefore, the data should be checked carefully throughout this version.
[Thank you. We have carefully checked the data throughout this version.]
Reviewer 2 Report
Some minor tips in main document
Comments for author File: Comments.pdf
Author Response
Thank you very much.
1) regarding your comments in the page 2, line 52, we have deleted the [22, 23] from the beginning of this sentence; it was a type mistake.
2) regarding your comments in the page 11 on figure 5, we have adjusted the label of the y-axis to make all texts fit in one line.
