Mapping the Spatial-Temporal Dynamics of Vegetation Response Lag to Drought in a Semi-Arid Region

Round 1

Reviewer 1 Report

Authors:

With pleasure I went through your fine and comprehensive study. Therefore, I have few comments only.

What puzzles me is that you select the period 2006 – 2009 as having “normal years” . However, 2006 had high temperature and low rainfall anomalies. Why include that year, or is it that with years 2007- 2009 some average conditions are obtained? Please comment on this in the text.

328, 457  DEM. You have not defined DEM (Dig. Elev. Models?) and not clear is how you have used DEM data in the analysis. It does not appear in fig.2.

Editorial:

Table 1: extent (pixels) 3500x1569  .. ?   What do you mean ?

175: … that not planted..  syntax error

226: .. SPI but represents in daily scale.  Should be: represent it in …

281: … DTW is an important for reducing..  Delete “an”

399 … irrigation few irrigation…. ?

495    random forest modelling

506/7 The (delete The) further studies

Author Response

We sincerely thank the reviewers and editors for their valuable comments that helped us to improve our manuscript. We have made revisions carefully by following the comments. Our responses to the reviewers’ comments are listed point by point as follows (please also check the attached response letter, of which the contents are exactly the same but is more convenient to read). 

       Please see the attachment please.

Author Response File: Author Response.doc

Reviewer 2 Report

Dear editor and authors,

Excuse my limited English. I think it is an interesting work for the study of water resources, and scientific quality is high. However, from the point of view of remote sensing it is not too relevant. The methods are not novel (they are standardized and widely used methods), although the use of all of them together gives the work an added value. I also believe that there are some aspects that should be improved. Because of this, I recommend that a major revision be made before considering its acceptance.

A brief summary of some recommendations that could be corrected.

a) GENERAL COMMENTS:

1. Introduction: I think the introduction is interesting and well-structured but I think it should be revised and completed. Many citations are missing, there are complete paragraphs with statements but without citations. In the specific comments I note some aspects to improve. In addition, I do not believe that the objective is reached: The paper does not discuss human activity in the response of vegetation.

2.  Materials and Methods: I think the section is well explained except for some specific annotations. I think that Figure 2 is very interesting (in this type of work I think it is mandatory). However, I believe that pre-processing should be included (e.g. sub-section 2.2.5). I think it is more logical that this figure and a summary of the process be included at the beginning of section 2.

2.1. Study Area: It's short but I think it's enough for the purpose of the study. However, the authors include information without citing the source.

2.2. Data and Preprocessing: I think it's well explained. However, the authors should comment on how they have homogenized the spatial data: in particular, how do they homogenize the different cell sizes? What resampling method? What is the final work resolution?

Section 2.2.1 .: the authors should explain the section better: they should comment if Nebraska is covered with a scene or several must be joined. If the data is subjected to a quality process, what dates have been used?

2.3.1. RF: I think it's a relevant section that should be expanded. In the specific comments I comment some aspects.

3.  Results and Discussion: I think the section is well explained. I liked the simplicity in the explanations. However, there is no discussion. Authors should review the section and include reference to other studies as a discussion.

b) SPECIFIC COMMENTS:

1. Introduction

C.1.- Line 69. What studies? There are only two studies and the conclusions are not commented.

C.2.- Many citations are missing, there are complete paragraphs without citations, for example:

·         The paragraph that begins on line 81 must be cited.

·         The paragraph that begins on line 92 should be reviewed and include citations.

C.3.- Should be cited similar studies to this work, I think the introduction should be developed a little more; should be oriented to a reader who, without being a specialist, understands the context of the study (similar studies where SPI with NDVI is used together, etc.). Authors should compare their methodological approach with other approaches used in different works. In this sense, I think it is interesting that the authors comment on some works with SPI and why they select the Gamma function and not Pearson III or log-normal.

2.  Materials and Methods

2.1. Study Area

C.4.- Figure 1: Please, add reference or data source in map (CDL). In addition, it must be completed. I also think that it should be completed, or create a new figure with the study area (for example, reference map, states, cities, etc.); readers who do not know the area should be located in it. Add units in scale bar in International System (in all figures).

C.5.- Table 1 is not necessary, the values can be included within the text (and cite its origin).

C.6.- Given the objective of the paper, and the many doubts that arise, I think it is essential to include two figures of the spatial distribution of temperatures and precipitation to represent the spatial pattern.

2.2. Data and Preprocessing

C.7.- Line 143, 153, 155. Add resource in bibliography (NASA, RMT, LDOPE). The link http://wist.echo.nasa.gov/ does not work. On the web page of these resources it appears how to cite them (for example, RMT).

C.8.- Lines 143-144. Please add atmospheric correction algorithm and its reference. It is important because this article addresses a multi-temporal study.

C.9.- Line 144, Add cite to MOD09GQ:  Vermote, E., Wolfe, R. (2015). MOD09GQ MODIS/Terra Surface Reflectance Daily L2G Global 250m SIN Grid V006 [Data set]. NASA EOSDIS Land Processes DAAC. doi: 10.5067/MODIS/MOD09GQ.006

C.10.- Line 163. Check the wavelength intervals.

C.11.- Line 166. Add resource in bibliography (ORNL DAAC).

C.12.- Line 167. Parameters or variables?

C.13.- Line 168-169. The authors say “The daily average temperature and precipitation over Nebraska were selected”. I do not understand the phrase. Do they work with daily average data with temperature and precipitation? Please, explain it better.

C.14.- Line 173. Add resource in bibliography.

 C.15.- Line 181. Add resource in bibliography.

C.16.- Line 186. This subsection should go after 2.2.1.

C.13.- Line 191. Refer Savitzky-Golay filter.

2.3. Methodology

C.14.- Figure 2. It should also be included Irrigation Data and its use. All processes, including RF, should be included.

C.15.- Line 222-229. Why first an annual SPI to identify the drought and then use monthly SPI? Why not analyze all dry periods from monthly SPI?

C.16.- Eq. 3. Check, is tau(aplha)

C.17.- Eq. 5-6. Refer approximation. Check constants notation (C0, C1 and C3).

C.18.- Lines 253-267. I have some doubts about the temperature anomaly, I think they should clarify it: “The difference between the observed value and the multiyear Gaussian fitting value is used as the index of the climatic anomaly (Equation 9)” but Eq. 9 is a difference between the observed value and the average value of the values. The average can be adjusted to a normal curve but it is not a perfect normal. The paragraph is somewhat confusing, I think it is better to rewrite the entire paragraph so that it is clear. Also, equation 9 and its explanation should go where it is referenced, not after Figure 3.

C.19.- Eq. 9. They must define t, its day? Is the equation applied to each pixel, by zones or the adjusted normal curve is unique to all of Nebraska?

C.20.- Figure 3. Define the caption (b). ¿What is DOY? In the text, subfigures (a) and (b) should be commented.

C.21.- Line 273. For consistency with the position of the figures, better “Temperature observations and Multi-year SPI and in Nebraska ….”. Also, did you use temperature observations or temperature anomalies?

C.22.- Subsection 2.3.4. I think it's very interesting. My question is: how do you get the best response lag?

C.23.- Subsection 2.3.5. Please refer RF (Breiman 2001). This section should be developed, I think it is important. Some aspects should be described:

·         They must justify the choice of covariates. They must also first analyze the collinearity between variables; In the case of RF it is important.

·         What kind of observations have you used? (Example pixel, etc.)

·         How do you get the value of the parameters mtry and ntree? Have you calibrated the model?

·         RF must be validated because it tends to overfit the model. Authors should do a validation or cross-validation

·         What index of importance is used? The effects should also be estimated

C.24.- Figure 4. I think it should go after being referred, not before. The same thing happens with other figures, please check.

C.25.- L.357. “The occurrence of scarce rainfall has no spatial rules”. I do not agree. It is true that rainfall is very complex compared to other variables, and it is difficult to predict, but it is very dependent on local spatial effects. These are difficult to include in the models but they exist (for example, the generation of specific events by the orography). The problem is that it is very difficult to include local factors (which are generally spatial) and difficult to predict.

C.26.- Figure 6. Is it part of the results?

C.27.- Subsection 3.2. In this section, a discussion is especially important. I would like to ask a question about the estimation of thermal anomalies: is it possible that when using a single normal curve it could affect the results?

C.28.- Subsection 3.3. They must complete it with references (discussion).

C.29.- Subsection 3.4. I still believe that there is a problem in the anomaly of the temperature derived from how it was obtained.

C.30.- Figure 9. What importance index? Define which index has been used as importance.

C.31.- Line 460. What studies? They must relate assertions they make in a discussion.

C.32.- Conclusions. Please, do not include in conclusions a numbered list, better include as paragraphs.

Author Response

We sincerely thank the reviewers and editors for their valuable comments that helped us to improve our manuscript. We have made revisions carefully by following the comments. Our responses to the reviewers’ comments are listed point by point as follows (please also check the attached response letter, of which the contents are exactly the same but is more convenient to read).

       Please see the attachment.

Author Response File: Author Response.doc

Reviewer 3 Report

Even if this paper has good potentialities, you only partially deal with the topics. Both methodological and results chapters must be detailed and reorganized.

Some eco-physiology and climatology concepts should be improved.

English needs to be revised and the bibliography has to be integrated.

 

Below, some recommendations and corrections:

1.     Line 42: Maybe you would write “insufficient rainfall” instead of “sufficient rainfall”?

2.     Lines 69-75: clarify and simplify the sentences using better terminology. Anomalies of the growing seasons highlighted by the VI along years can be due to several reasons, not only to extreme weather events like droughts, but also to agricultural practices as i.e. crop rotations or different planting times (the correct words are “winter crops vs summer crops”, not “cool versus warm season vegetation types”).

3.     Lines 97-103: You should include some references to strengthen your assertion, e.g. Van Hoek M. et al., 2016, https://doi.org/10.3390/rs8050422; Wang J. et al., 2003, DOI: 10.1080/01431160210154812; etc. Did you consider the hypothesis to use the EDI-Effective Drought Index, that is a daily index, to overcome the problem of the monthly time scale?

4.     Lines 132-134: Could you include some reference confirming the sentences?

5.     Line 168: Use “mean temperature” instead of “average temperature” to indicate an average between the daily max and min temperature.

6.     Lines 174-178: Did you use CDL to locate crops only during drought periods? Could reformulate the sentence? Check the grammar of the sentence and simplify it.

7.     Lines 183-184: Why do you use only 2012 MIrAD data?

8.     Paragraph 2.2.5.: What are the results of the filtering? Did you obtain a good, continuous time series? Show some statistics, or comment your results defining the quality of the NDVI data set.

9.     Paragraph 2.3.1.: The description of the SPI calculation is unnecessary; It’s sufficient to report the main reference. Instead, I suggest defining what is the threshold that you considered in order to identify a drought event, i.e. all SPI values below -1 or other thresholds? Explain your choice.

10.  Line 224: What do you mean with “extreme drought year”? Based on the SPI classification the words “Extreme drought” specifically means values below -2.

11.  Paragraph 2.3.2.: You have to classify the temperature anomaly into “negative”, “normal”, “positive” in order to simplify the comprehension of the maps, choices, and results.

12.  Lines 255, 258 and 261: see the comment of Line 168.

13.  Lines 273-274: Did you mean Figures 4 and 5 instead of 3 and 4? Moreover, the first time you describe figures, you have to arrange them as near as possible to the sentence. The SPI maps have to be classified in order to better identify areas and periods affected by droughts.

14.  Lines 274-276: Looking at the maps is not clear how you identify the 2006-2009 as “normal climate” period. You have to clarify if you consider both SPI and temperature anomaly and which weight you attribute them. The classification of SPI and temperature anomaly facilitate comprehension.

15.  Line 277: “…by averaging the daily NDVI values…” instead of “…by averaging the NDVI values…”

16.  Line 278: “of” instead of “or”

17.  Lines 278-280: “These curves represent the mean growing season NDVI of each vegetation type under mean climate conditions.” instead of “These curves represent a typical vegetation-growing phenological response of each vegetation type under normal weather conditions.”  Using “normal” word imply that you averaged the NDVI only considering years that previously you declared as “normal climate years”. Moreover, graphs of the mean growing season NDVI for main crops and grassland could help to understand the vegetation cycle under mean climate condition.

18.  Lines 281-282: growing season NDVI anomalies could be also due to crop rotation (cultivation of different types of crops in the same area in sequenced seasons to maintain soil fertility).

19.  Lines 287-292: A graphical example of the application of the DTW should be included to better explain how the procedure works.

20.  Lines 296-298: Could the fluctuation of the NDVI profile during “non-growing season” be due to crop rotation (e.g. winter crop-summer crop)? You should verify it. What do you mean with “non-vegetated background signal”?

21.  Line 300: “start and end of the growing season” instead of “start and end of growth in phenological profile”.

22.  Lines 314-316: Why you use first “precipitation/temperature” and then “SPI/temperature”? What are the “different response scales”? A better explanation is needed.

23.  Lines 348-349: Specify that Figure 6 represents a comprehensive drought index including a series of indices; include the legend into the graph.

24.  Lines 349-350: You declare that in 2000, 2003, 2006 and 2014 drought was more localized, but if we consider all the intensity classes, 70-80% of the State was affected by drought. So, what classes do you consider in this study? There is a correspondence between U.S. Drought Monitor comprehensive index and SPI 12?

25.  Lines 350-353: The greater irregularity of rainfall is not limited to the growing season, but is due to several factors, e.g. the atmospheric circulation, the morphology of the territory, etc. you need to reformulate the sentence.

26.  Line 355: “the only arid disaster area” is a strong sentence, but if we look at the SPI 12 legend -1.3, that correspond to moderate drought, is the highest value reached. Moreover, you have to pay attention when you use the term arid as synonymous of drought because they have a different meaning. (https://www.ncdc.noaa.gov/monitoring-references/dyk/drought-aridity)

27.  Lines 357-359: What does it mean that drought is vulnerable to the interference of heavy rainfall events? Not always a heavy rain ends a drought period. Reformulate the sentence.

28.  Lines 363-364: Do you have references that support this concept?

29.  Line 368: “dryness” instead of “desiccation”.

30.  Lines 374-375: Redundant sentences. They should be simplified.

31.  Lines 392-394: clarify the concept. During normal or wet climate condition, vegetation should not have response lag, but its response along an accumulated rainfall deficit that follows a normal period is slower due to the water availability in the soil.

32.  Lines 400-403: The only use of remote sensing Vis to detect and explain the different physiological response of vegetation to drought conditions in agricultural lands must be carefully considered. For example, another possible explanation of the stronger response to drought in the western part of Nebraska is that crops of the eastern part are irrigated (corn, in particular, is a crop with high water requirement). For this reason, yearly irrigation scheduling should be considered, especially during drier seasons in order to compare different behaviors.

33.  Lines 404-410:

34.  Lines 420-422: the lag period of at least 2 weeks is also due to the accumulated soil moisture. This result could also imply the possibility to use an SPI 1 with a moving window of 1 week instead of 1 day.

35.  Line 424: specify: “response lag frequency distribution”

36.  Lines 435-439: Trees have a longer response lag to prolonged drought because of their deep root system.

37.  Paragraph 3.4.: Random forest should be the first analysis described at the beginning of chapter 3 – Results and discussion, because it supports the other results, but it also suggests constraints that have to be considered when an analysis of vegetation response to extreme events is made.

38.  Lines 459-460: Vegetation is influenced by heat waves (temperature above normal lasting from a few days to a few weeks) more than a daily temperature anomaly.

39.  Lines 482-485: point number 2 is not a conclusion. It is a normal pattern of precipitation compared to temperature.

Author Response

We sincerely thank the reviewers and editors for their valuable comments that helped us to improve our manuscript. We have made revisions carefully by following the comments. Our responses to the reviewers’ comments are listed point by point as follows (please also check the attached response letter, of which the contents are exactly the same but is more convenient to read).

Please see the attachment.

Author Response File: Author Response.doc

Round 2

Reviewer 2 Report

I would like to congratulate the authors, I think the study has improved a lot and is very interesting.  I also believe that they have made a great effort and all my suggestions have been included. Because of this, I believe that the document can be accepted.

However, I would like to answer Response 26 of the authors (in red) and suggest that they include some of these aspects (as a suggestion):

Response 26: For first question, in our view, as an ensemble learning method with multiple decision trees, the main advantages of RF include significantly reducing possible overfitting (Patel et al., 2015), and little influence by collinearity among analysis variables (Cutler et al., 2007). Collinearity analysis between variables is not important in RF analysis rather than other machine learning method.

RE: Maybe I expressed myself wrong, sorry for the confusion. I agree with Cutler et al. (2007), pore so me gusta mucho RF. That's why I like RF a lot.

RF works very well in prediction when there is collinearity. The problem lies in the importance score in the variables (and the interpretability of the model). For example, if there are two highly correlated variables, the model will use any of the two without a clear preference and when assessing the importance it will not really be known what happens. Conditional permutation importance can solve this problem in a way. There are several works on this subject; for example Gregorutti, Michel, & Saint-Pierre (2017).

It is not necessary to recalculate the model (there are few variables), but I think they should evaluate the collinearity; It is simple (correlation, VIF, etc.) and can lead to a better interpretation of Figure 9.

For third question, we used matlab to run the random forest, unlike the R package, RF in matlab have two parameters need to be tuned : ‘MinLeafSize’ and. ‘NumTrees’, ‘MinLeafSize’ is the minimum number of observations of per tree leaf, Default values is 1 for classification and 5 for regression. The optimal ‘MinLeafSize’ was decided by comparing mean squared errors obtained by regression for various leaf sizes using 1/3 of dataset. ‘NumTrees’ was then selected by the “out-of-bag squared error”(oobError) (https://www.mathworks.com/help/stats/regression-treeBagger-examples.htm).

RE: Ok, it is the TreeBagger function of MATLAB. RF has the parameters mtree and ntry. In TreeBagger mtry is NVarToSample. In case of regression, mtry=(1/3*nfeature). However, you can refer that one of the advantages of RF is that with the default parameters the differences are not significant with respect to calibrated models. Example: Gomariz-Castillo, Alonso-Sarría, & Cánovas-García (2017).

For forth question, treeBagger class in matlab is a Bootstrap-aggregated (bagged) decision trees combine the results of many decision trees, which reduces the effects of overfitting and improves generalization. So the model is validated by the out-of-bag-error in the matlab code.

RE: I agree. However, all models must be evaluated with a validation or cross-validation (must be evaluated with data not used in training), although it is true that out-of-bag-error is a measure of RF error.

However, since the objective of RF in this study is to evaluate the importance of the variables, not their extrapolation capacity (it would be another issue to extrapolate outside the variable space), I believe an estimate of the model error would be sufficient.

Bibliography:

Gomariz-Castillo, F., Alonso-Sarría, F., & Cánovas-García, F. (2017). Improving Classification Accuracy of Multi-Temporal Landsat Images by Assessing the Use of Different Algorithms, Textural and Ancillary Information for a Mediterranean Semiarid Area from 2000 to 2015. Remote Sensing, 9(10), 1058. https://doi.org/10.3390/rs9101058

Gregorutti, B., Michel, B., & Saint-Pierre, P. (2017). Correlation and variable importance in random forests. Statistics and Computing, 27(3), 659–678. https://doi.org/10.1007/s11222-016-9646-1

Reviewer 3 Report

Authors have answered adequately to my comments.