Quantifying Forest Fire and Post-Fire Vegetation Recovery in the Daxin’anling Area of Northeastern China Using Landsat Time-Series Data and Machine Learning

Round 1

Reviewer 1 Report

Overall, authors present an interesting approach to map and assess disturbances especially those related to wildfires in the study region. Despite these positive aspects, the manuscript presents many shortcomings, namely (see detailed in-line comments in the ‘specific comments’):

• A low number of sample sites to perform validation (n=3). Despite these sites are relatively large (3x3 Km areas) thus generating a high number of pixels, there is no guarantee that these three sites cover the heterogeneity and diversity of spatial, environmental and spectral conditions in the area;
• Several methodological options and parameters are not fully detailed in the text;
• The use of NDVI at year five after the fire (response variable) in modelling lacks to present robust criteria for its selection (why year five exactly?) and fails to contrast it to pre-fire conditions. As such, this variable seems inappropriate to capture the recovery trajectory and hence unable to fully depict the return to pre-fire conditions in both speed and completeness terms;
• The discussion section is often out-of-focus and speculative especially regarding the evaluation of post-fire recovery actions, forest management practices or policies which the manuscript data and analyses do not support.

 

In general, some effort to revise and further improve the overall flow of the text, some grammar issues and correctness are also needed.

 

Specific comments

Introduction

L17: It is unclear how post-fire vegetation recovery will aid to respond to climate change. I believe there are local effects that will benefit more directly from this approach including the re-establishment of ecological functions, species habitat recovery, biota recolonization, erosion control, etc. In this regard, I advise the authors to reflect more on this issue.

L40-41: “(…) the losses evaluation over time for different regeneration intervals.” – this sentence is strange and requires clarification.

L42-L43: The statement along these lines requires better support from previous literature to establish the importance of fire as a driver of local/regional scale climate effects.

L44: the reference on [9] mentions the contribution of the disturbance-recovery dynamics to carbon budget not the way vegetation recovery influences climate…

L54: seems that a word is missing in this sentence: “which has been extensively used in (…)”. Please check and correct.

L54-55: This phrase needs clarification and better context. Moderate-resolution imagery may be more appropriate but it strongly depends on the objectives of the study. For some applications high/very-high spatial resolution is in fact more suited (e.g., structural damage assessments). Also, other moderate resolution platforms such as Terra/MODIS are often more tailored to perform wildfire assessments given their very-high temporal resolution (~daily vs 16-day of Landsat).

L61: A short and ”top-level” description of the VCT method here would help the reader to understand better how it works. Consider adding it.

L63-63: Needs clarification. Do you mean “frequently managed forest areas”?

L70: Reference in 33 lacks square brackets.

L78: Needs space after the stop.

L79: Correct to “a critical”.

 L83-84: The sentence needs rephrasing and clarification as its meaning is unclear especially around here: “… one of the most critical forestry based …”

 

 

Methods

L92: Indicate the size of the study area (in Km2 for instance).

L98-99: the contents of this report are unclear… why is this relevant to the objectives of the manuscript?

L144: Correct to “semi-variogram”.

L145: Was the exponential function used for all climatic variables (i.e., temperature, precipitation, moisture)? The reported R2 respects to which variable exactly?

L147: Instead of “histories” perhaps “time series” or “chronology” would be a more suitable term.

L149: Square bracket lacks in reference 22.

Disturbance assessment uses the VCT algorithm but the source of forest maps (supposedly used to mask/select forest pixels) is not defined. Further, in line L161 forest data discriminates between sparse and dense forests which have different threshold values. How was this forest data obtained or its source? Or, is the VCT method responsible to classify spectral data into forest/non-forest with different densities? How is this achieved? Overall, this needs much further clarification.

L161: Briefly describe how threshold values were obtained. Which statistical analysis was employed?

L162: “The consecutive high IFZ values were used to avoid noise problem.” – what is the meaning of this and how is it relevant for the workflow? Have authors employed any filtering procedure to reduce noise? How was this performed? Please clarify.

L174: Replace by a more formal notation: “3×3 Km sample areas” instead of using the asterisk. Please clarify.

 

Only three sample sites used to evaluate the VCT forest disturbance products in such a large and heterogeneous area seems insufficient. Did authors check if this number of sites is enough to obtain robust estimates of accuracy/SAI (with low overall variance)? Also, I wonder if spatial, environmental and spectral variability is adequately captured by these three sites?

I would recommend to check this issue carefully and increase sample size if deemed necessary by tests. One way to check this would be to plot the distribution (e.g. boxplot) of spectral variability values per plot (e.g. standard-deviation) across different sample sizes (e.g. 3, 5, 10, 20) and assess when the median starts to asymptotically “stabilize”. This is equivalent to species accumulation curves in ecology which are used to assess if a survey can capture the species pool or if more samples are required. For example, the work by Rocchinni et al (2009) unifies both spectral and species curves for assessment.

Rocchini, D., Ricotta, C., Chiarucci, A., de Dominicis, V., Cirillo, I., Maccherini, S., 2009. Relating spectral and species diversity through rarefaction curves. International Journal of Remote Sensing 30, 2705-2711.

 

L190-194: Authors have used several features based on spectral bands and indices in SVM-based image classification (subsection 2.3) which arises several important issues:

• Was collinearity among features tested? Were highly correlated variables removed before performing the SVM training? EVI and NDVI for instance are highly correlated indices that would make sense to select only one. This issue is addressed below for modelling post-fire NDVI but not here, any reason for this?
• Which dates were used to extract/calculate features and train the SVM? Were features calculated with images solely for the post-fire situation or were these extracted for both pre- and post-fire conditions? This needs clarification.

L194: Which components of the TCT were used as features? Do state these for clarity sake.

L199: I advise authors to state exactly (at least in supplementary information) how many pixels per year were used as labelled train examples including both burnt and unburnt pixels.

L200: Correct stop punctuation.

L204-205: Holdout or split set cross-validation was used to assess the classification performance of SVM but the number of rounds used to randomize the train/test splits is not defined by authors. Several rounds (n= 20, 30, 50, 100, …) are often used to estimate performance measures and then averaged. Please clarify.

L207-208: how was the gamma parameter defined? Which value was used? Saying it was parametrized is rather non-informative…

L224-226: A couple of issues here regarding these modelling procedures, namely:

• Which criteria were followed to select the five-year NDVI? Why not select other periods as well to better understanding medium and possibly long-term recovery (e.g. 10, 15 or 20 years after the fire) especially given there is a 45-year long time series;
• Instead of ‘directly’ using the NDVI value 5-years after the fire it would be better if the response variable was defined as the post-fire anomaly i.e. the difference in NDVI between the pre-fire average (as a referential) and the 5-years post-fire NDVI. That way the (spectral) recovery process would be better captured as the return to previous conditions thus nearing a value of zero anomaly as “full” recovery over time. By using the NDVI value “directly” authors are simply aiming to explain the amount of greenness at that particular point in time (i.e., state variable) without any reference to pre-fire conditions which has strong consequences in the analysis. This may explain why elevation (“dem”) is so important (Figure A1) since it may capture the elevation gradient in vegetation/biomass amount. Please comment on this issue and provide changes in the analysis pipeline if needed or relevant.

L233: replace by “algorithms”.

L275: Correct the reference to zhangwe paper.

L274-280: Again, holdout cross-validation is used but the number of evaluation rounds is not stated. Exactly how many train/test splits were performed?

 

Results

L288-289: It would be more informative to state exactly which years have recorded no disturbance in sample sites.

L321: Replace by "except for the year 2016".

L323-325: The aim of using the combination of VCT and SVM algorithms is to improve detection of burnt areas but then authors attribute this to using a two-class response and the limitation to >1 ha fires (which btw should have been previously stated in methods…). Overall, this sentence needs clarification.

L326: Why have ridgelines such similar characteristics? Fire scars typically have very distinctive spectral signatures (e.g., with much lower albedo shortly after the fire due to charred surfaces). Please elaborate on this issue.

L348: Replace by “RF model”?

 

Discussion

This section seems often a bit vague, speculative and away from the topic.

 

L352: this opening line is quirky… are you referring to time series of remotely sensed spectral indices?

L360: How were "invested recovery efforts" assessed in this analysis? This sort of statement should be avoided if it cannot be robustly supported by the data and analyses in the manuscript.

L368-375: These lines should be presented in the results section. Also, it would be better for the reader and to the overall flow of the manuscript to present Figure A1 in the main text instead of placing these important results in the Appendix. Also, Figure A1 caption requires more detail to describe the name of variables and be self-explanatory.

L384: I suspect dNBR (severity) is ranked low in importance because authors are aiming to explain NDVI values as a state variable without any reference to pre-fire average conditions (i.e. anomalies in NDVI). As such, NDVI_1 is substantially more important because it presents more temporal correlation to NDVI_5. Overall, there are much better approaches to capture recovery through spectral indices than the one used here (e.g. Bastos et al (2011), João et al (2018), Veraverbeke et al (2011)). As it stands, using NDVI_5 as a response variable will mainly depict which environmental factors affect the amount of greenness at a certain time (i.e. state) including the previous state (i.e. NDVI_1) and main elevation gradients (i.e. dem).

 

Bastos, A., Gouveia, C.M., DaCamara, C.C., Trigo, R.M., 2011. Modelling post-fire vegetation recovery in Portugal. Biogeosciences 8, 3593-3607.

João, T., João, G., Bruno, M., João, H., 2018. Indicator-based assessment of post-fire recovery dynamics using satellite NDVI time-series. Ecological Indicators 89, 199-212.

Veraverbeke, S., Lhermitte, S., Verstraeten, W.W., Goossens, R., 2011. A time-integrated MODIS burn severity assessment using the multi-temporal differenced normalized burn ratio (dNBRMT). International Journal of Applied Earth Observation and Geoinformation 13, 52-58.

 

L404-407: The assessment and analyses on the manuscript were not tailored to answer questions related to the effectiveness of policies or management practices after the occurrence of wildfires. The comments are unsupported by data and references are highly speculative.

L413: Again, speculative and unsupported by valid scientific references or data from the study…

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This interesting manuscript used disturbance products from the Vegetation Change Tracker (VCT) algorithm and Support Vector Machine to generate historical maps of burned vegetation area. Authors then used three modelling approaches to determine the influence of environmental factors on vegetation recovery after fire disturbance. This is an interesting study exploring machine learning to investigate factors that affect forest recovery following wildfires.

Major comments

While this an important study for restoration ecology I have comments below that need to be addressed.

1. What was authors definition of recovery in this study? Why recovery based on five-years post-fire NDVI and how does this measure of NDVI correspond to various pre-fire vegetation types (trees, grass, shrubs)?

What was the relationship between the vegetation characteristics from the authors field work and this spectral index?

2. Line #91 – Has section heading “1. Data and preprocessing” but it is immediately followed by study area description. Split sections for study area description and data description. Maybe data description section can begin from line # 104.

Minor comments

1. Lines #28 – 30 The conclusion of the abstract should clearly state the most influential factors instead of a general statement that topography and climate affect vegetation recovery.
2. Line #176 – S_v is not part of equation 3; the mention of S_v is confusing
3. Lines #203 – 204 State the total number of points from which you developed training and validation datasets
4. Lines #208 – 210 ” … a pixel was considered as a fire disturbance pixel if it was labeled as adisturbance pixel by VCT algorithm and was also classified as a burned pixel by SVM simultaneously”

How did you ensure that non-fire disturbances happening prior to fire, in same year and location, were not wrongly labeled as fire disturbance simply because both disturbances occurred at same location

5. Lines #224 – 230 Authors briefly mentioned some of the input variables for the models. Consider showing all the input variables by repositioning Table A2 here. This should be a good place to let readers know the factors been considered.
6. Being an important result in this study, many readers would want to see the influential factors in the main manuscript. Authors should consider including the variable importance plots from the RF model in the Results instead of appendix.
7. In Discussion authors discussed meteorological factors as very important variables in the models; however, there is no results in the Results section to back their claim.
8. Lines #283 – 286 Under Results section should rather be part of the section on model validation under Methods
9. In Table 1 – What is the difference between column name “Year” and “Disturbance year”
10. Line #371 – What is preci_1? Similarly, what is NDVI_5 in line # 373, moisture_0 in line# 374. Please give full names. It is a distraction for readers to keep flipping pages to find their meaning.
11. Line #387 – 388 - …(French et al. 2008)… be consistent with citation format (annotation)
12. Line #399 – “supress” should spell suppress
13. Line #400 – “In addition, Forest” – forest
14. Line #403 – “conductive” – conducive?
15. Line #447 – “The RF, SVM and SMLR models both explained the…” Revise, you cannot refer to a list of three as “both”
16. In appendix, Table 1 A does not seem helpful other than making the appendix longer.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear authors,

This study investigates the recovery of vegetation in burned areas. Although the results were satisfactory, as the authors comment, more fixed-point field studies are needed to track the actual state of vegetation recovery to characterize the vegetation recovery processes.

But in general, the study is well presented and interesting for the readers.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Authors have provided detailed answers regarding almost all my previous suggestions and comments (especially those regarding methodological issues) and I would like to congratulate them for their efforts. However, in their rush to change the manuscript according to the provided feedback, I feel that some alterations were made without much care and attention resulting in not fully structured or matured ideas which impact overall text quality, flow and readability. This is especially evident in the discussion section which needs further revision to improve its overall quality and correctness.

Also, by not comparing the response variable (NDVI at year +5) to any pre-fire fire referential or baseline, it becomes difficult to estimate how much a certain area has recovered to its pre-fire state/condition (assuming here that the same state is maintained, i.e. no stable state shift). Given this limitation, which preferred course of action would be to do a reanalysis, authors opted to recognize this limitation in certain parts of the discussion but, as before, often there are bits of text that seem unconnected and need some improvement.

Regarding the relatively low sample size for evaluation (n=3), authors may also need to recognize limitations in their ability to capture the full diversity of conditions in the study area. In that sense, consider adding a comment on this issue in the discussion section along with tailored suggestions for improvement.

Reviewing the discussion more carefully may aid to solve some of these issues as well as comparing the main findings to prior research more thoroughly. One way of doing this would be to organize subsection 4.2 according to each factor importance (elevation, maximum temperature, NDVI +1, moisture, etc) and commenting/discussing each one in a more structured and connected fashion. Address not only the importance of each variable but also if the effect is either positive or negative (i.e. effect sign) to better understand how it influences recovery.

Take also into consideration the need to review and improve English language grammatical correctness along the text.

 

Detailed comments

Table 5 lacks a proper caption detailing the full name/description of variables with related acronyms. SLMR should also be in the extended format for clarity. Keep in mind tables should be self-sufficient and clear even without the need to access the full paper. Also, keep in mind the number of decimal plates/significant digits… does your method supports that amount of numeric precision? Please check for instance estimate, std error, etc. Also correct names at column #1 (e.g. mean_high_, mositure_0, mean_hig_4 …??)

Figure 5 – Increase figure and/or acronym names size for better visualization. Provide a contained description of acronyms for making the figure self-sufficient and more easily readable if outside its original context. By not having this information in the caption makes the reader go back and forth to understand the meaning of each acronym...

Figure 6 – Please avoid using acronyms in captions. These elements are meant to be self-sufficient and easily readable.

L28-29: Putting acronyms here without previously defining them is completely uninformative and should be completely avoided. Use names in extended format. Check other instances in the text where this may happen and perform replacements to improve text readability if necessary.

L339: The wording seems a bit strange. Maybe rephrase to: “(…) was one of the main limiting factors for vegetation recovery in this area.”?

L369: “fire recovery are the crucial period that shape forest age structure and forest type.” – fire recovery is a process, not a period?! What do you mean here? Please clarify and provide appropriate references

L371, L376: Do you mean “good performance” or similar?

L377, L379: Instead of using “other researches” maybe “previous research” seems a better option

L382-383: It is really hard to follow the meaning of this sentence. Please revise it

L389: “the validation performances of the models are inferior to those of the modeling,” – the wording here is really difficult to grasp. Are authors comparing results of model performance between fitting and validation data? Please clarify and correct

L394: “the machine learning models of RF and SVM” seems quite strange wording. Maybe “machine learning models such as …”?

L397: A really strange sub-section name… (probably as a result of automatic translation?) Maybe: “Limitations and future improvements” or something similar? Please check.

L400: Instead of “monitored” perhaps change to detected?

L404: Perhaps “restoration” is not the best word here as it implies any sort of action. Maybe best to use recovery or recuperation?

L411, L417: Landscape is not a scale… It is a level of biological organization which complexity derives from multiple sorts of interactions between abiotic and biotic components. Please check proper terminology in landscape ecology domain.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Authors have done well responding to all comments. However, authors can do a better job incorporating their responses to the major comments in the manuscript. They have offered their explanation to me as a reviewer, it’s time to do same for their readers.

Line # 28 – 29: The use of abbreviations like preci_2, moisture_5, and ex_high_3 in your abstract will only confuse your readers. Find a better way to mention these variables.

Author Response

Please see the attachment.

Author Response File: Author Response.docx