Monitoring Landscape Dynamics in Central U.S. Grasslands with Harmonized Landsat-8 and Sentinel-2 Time Series Data

Round 1

Reviewer 1 Report

This paper describes tests whether harmonized Landsat/Sentinel data can be used for tracking seasonal dynamics in grassland ecosystems. 

I think the general goal of the paper is laudable and should eventually be published. However, I think a few key components need to be dealt with first.  To me, the main issue is that the paper seems to waffle on whether the tests herein are simply to determine if the HLS data improvesthings relative to Landsat 8, or whether HLS-based analysis is sufficient to capture the underlying processes.  See my detailed notes for some contradictions in the text.  By and large, the spatial patterns of phenology seem to be driven entirely by the frequency of  the HLS or L8 data, and while the authors clearly acknowledge this, my guess is that most readers would want some guidance on whether it even makes sense to use Landsat/Sentinel data to monitor these things. At its core, it seems the authors are arguing that grassland dynamics happen at spatial scales too fine to be well-tracked with MODIS-scale pixels, and yet once we get into the results, the authors seem to shy away from the specificity of spatial pattern. 

Another theme of my comments is that the logic of the workflow is sometimes confusingly described, or simply under-described.  I have specific comments on where it seems it could be improved.   The figures and graphs could use some improvements I’ve noted. 

Finally, there should be some argument for why these particular case studies were chosen, out of all of the grassland themes that exist.  The paper feels like a hodge-podge of somewhat unrelated questions; I’m okay with this in a general sense, but it requires a bit more logic or argument. 

Detailed notes

Lines 50-51 – I’d suggest adding semicolons or rephrasing the clauses a bit more clearly – right now, it’s unclear if the food and fuel to cattle is the end of a clause, and if wild ungulates are provided by the habitat, or if the food and fuel is for the ungulates. Best not to trip up readers on the first line!  

Lines 78-82 seem to belong more clearly with the paragraph that starts at line 88

Questions on lines 84-87 appear abruptly – need more context for why these were chosen.  If this is driven by the HLS tiles, that’s fine, or by prior studies, etc., all good, but there should be some rationale. 

The introduction’s argument is generally fine. However, the motivating questions appear fairly abruptly (lines 84-87), and the specific  phenological questions of interest do not have clear context.  I suggest a bit of expansion  to frame those questions more clearly – why choose these, and why not others?    

General argument for introduction is good, though – an intriguing study. 

105:  For non-specialist readers, it maybe good to describe briefly what the HLS tiles are, and why you used 2016.  

106:  km2 should have the 2 superscripted

114:  to aid in your argument for using moderate resolution data, it would be useful to estimate the approximate footprint captured by the phenocam – this would illustrate why a MODIS-scale pixel may not be appropriate. 

 118:  superscript on km2

124:  what is the BAECV data?  Please provide a bit on the spatial scale, sourcing, etc. 

Figure 1:  scale bars on bottom two sub-figures are hard to read. 

Personal preference, but I find the legend took a bit more effort to interpret than it should – it is not immediately apparent that the legend applies to all three subsets, so I was looking around in the upper right and lower left figure for what the dots represented;  Perhaps just use arrows and words on the figures for the dots on those two, and eliminate them from the legend?   

Figure 2 – this makes sense. 

Section 3.2.1:  Either here or above where I noted, it would be good to indicate the spatial footprint of the phenocam, to get some sense of the representativeness of the data relative to pixels. 

3.2.2  Seems the title of this section should be different than the prior

Given the argument for the paper – that we need to see if moderate resolution data can get phenology -- it’s confusing to me that we would be testing it against a dataset that is somehow influenced by MODIS data.   More justification in this section would be helpful. 

3.2.3  This section is confusingly written.  The expectation I had leading up to this section and 3.2.2 was that we’d be comparing the HLS results directly to the phenocam;  thus, the introduction of 250-m based analysis is odd.  I think I can infer what’s going on, but that’s partly guessing.  Also, the specifics of the workflow in this section are unclear.  So, how is the comparison actually happening?   We need that logic first, but it’s not really stated anywhere yet. Then we can be informed about the need to only do that at pixels dominated by grassland, and finally how you built the mask for that purpose. 

3.4:   It is not clear to me how precipitation 34km away from the wetland site will be used for reference data – more explanation here would be helpful.  This is particularly necessary because the argument of the paper for moderate resolution information would seem – at first glance, at least – to be negated by comparing against distant data.  The whole wetland area would necessarily be treated the same, right?  In that case, couldn’t we just use a coarser resolution sensor?  I suspect not, but more logic to bolster your argument would help. 

Line 240:  suggest “temporally normalized using a compositing strategy described in the steps below:” 

Line 245:  It takes a leap of imagination to catch the fact that this is the compositing strategy that results in making the time step the same – clarify this a bit more.  

Step 7:  this seems to describe the comparison with the eMODIS data, but not the phenocam.  Was just a single 30m pixel used to compare with Phenocam?  

272:  Add in which component of CCDC’s outputs you actually considered to be the results – presumably the disturbance date, but please specify. 

Figure 3:  I appreciate the authors being honest in showing these results, but they certainly do not cast the moderate resolution data in a very favorable light.  At first glance, it appears the eMODIS data are capturing variability in spatial patterns commensurate with the likely spatial patterns of the fields, and that the temporal frequency drives all of the pattern in the moderate resolution products.  

333:  This is confusingly written, I think – do you mean that you took the 250-m grassland cells, identified all of the 60+ Landsat pixels within that and calculated the stdev of SOST across those Landsat pixels? Figure 5 would suggest that, but I’m not clear.  

Figure 4:  I appreciate the argument about there being a tighter distribution.  However, the motivation for the paper appears to me to be very spatial-scale-specific – can moderate resolution data do better at getting spatial patterns without the problems of mixed pixels that occurs with MODIS-scale pixels?   To answer that – especially given the patterns in figure 3 – we would want to see a scatter plot of MODIS SOST vs mean(median) SOST from corresponding cells in HLS data, not just the overall regional-scale means and distributions.  Just limiting it to the grassland pixels is fine, since the map in Figure 3 presumably contains a bunch of non-grassland pixels that confuse things.   Maybe the grassland cells do compare well on a pixel-by-pixel basis?  Upshot: I don’t think that these boxplots are the strongest arguments for this approach.   They can be kept in, but pixel-to-pixel comparisons seem more appropriate for the argument of the paper.

348 : typo on Vierling’s name

356:  nitpicky, I realize, but the HLS time series timingcorresponded well, but the NDVI absolute values were quite different.  Since timing is what you’re after, this is fine, but it would help to be precise in the language here since this is a central finding. 

Figure 6—please make the font size of the legend larger – there is room on the graph, and right now it’s quite small; same for the x-axis. 

The finding of the figure is encouraging, though!!!  

Figure 7 – please put scale bars on these.  Also, for subpart c) indicate that the reference products are MTBS and BAECV – the figure should stand alone without reference back to the main text.  

It would be instructive to see some actual pixel-level trajectories to see why the early season fires were detectable using individual Landsat images for MTBS and BAECV, but not the time series.   

390:  water surface elevation data?   These are not mentioned earlier, but they seem to be a critical reference dataset that would alleviate my concerns about only validating relative to a distant weather station. 

Figure 8 – this is a useful figure! 

423-431 – this is a good summary of what was found. 

519-521—Understood, but it would be useful to comment on whether it makes sense to use these data vs MODIS data, since that is how the introduction seems to set up our expectations.  And to be clear, you do state at the beginning of the conclusion section that you’re testing whether the Landsat-8 plus Sentinel provides “sufficient frequency … to enable monitoring seasonal dynamics…” – that suggests that you’re not just testing whether there is improvement, but whether the underlying processes can be tracked.  

Author Response

We would like to thank the reviewers for their constructive comments and suggestions. We have revised the manuscript accordingly. Please find attached a point-by-point response to reviewer’s concerns. We hope that you find our responses satisfactory and that the manuscript is now acceptable for publication.

Reviewer 1:

Lines 50-51 – I’d suggest adding semicolons or rephrasing the clauses a bit more clearly – right now, it’s unclear if the food and fuel to cattle is the end of a clause, and if wild ungulates are provided by the habitat, or if the food and fuel is for the ungulates. Best not to trip up readers on the first line! 

Accepted. The sentence now reads

The U.S. Great Plains provide habitats of global significance to migratory volant species [1-3]; important sources of food and fuel to cattle, wild ungulate species such as deer, and people; and serve a role in storing organic carbon [1, 4-5].

Lines 78-82 seem to belong more clearly with the paragraph that starts at line 88

The paragraph describing our research questions was modified to provide more explanation for our study. We hope these changes will resolve this clarity issue.

Questions on lines 84-87 appear abruptly – need more context for why these were chosen.  If this is driven by the HLS tiles, that’s fine, or by prior studies, etc., all good, but there should be some rationale.

Sites were chosen based on data availability and certain characteristics such as Phenocam data or fire history. We have modified the text to provide context for the selection of our sites (line 85 - 87). We also added text to the preceding sentences to describe the challenge of monitoring these features with coarser-resolution data (line 84 - 85).

The introduction’s argument is generally fine. However, the motivating questions appear fairly abruptly (lines 84-87), and the specific phenological questions of interest do not have clear context.  I suggest a bit of expansion to frame those questions more clearly – why choose these, and why not others?   

Thank you for pointing this out. More explanation was included within this paragraph to better introduce the issues and site selection. We modified the text to add clarity regarding our research goals.

General argument for introduction is good, though – an intriguing study.

Thank you. We thought the results from testing the HLS would be of interest to the remote sensing community.

105:  For non-specialist readers, it maybe good to describe briefly what the HLS tiles are, and why you used 2016. 

Accept and added. (line 143 - 146)

106:  km2 should have the 2 superscripted

Changed. (Now line 109)

114:  to aid in your argument for using moderate resolution data, it would be useful to estimate the approximate footprint captured by the phenocam – this would illustrate why a MODIS-scale pixel may not be appropriate.

Accept and added: “The footprint of PhenoCam is about 11 Landsat-8 pixels that forms a sector shape within 5 by 3 pixels and covers one MODIS 250m pixel.” (line 188 - 190)

118:  superscript on km2

Accept and changed (line 121 and 132).

124:  what is the BAECV data?  Please provide a bit on the spatial scale, sourcing, etc.

Accept and added 30-m and also a reference for the BAECV data product description (line 127 - 128).

Figure 1:  scale bars on bottom two sub-figures are hard to read.

Personal preference, but I find the legend took a bit more effort to interpret than it should – it is not immediately apparent that the legend applies to all three subsets, so I was looking around in the upper right and lower left figure for what the dots represented;  Perhaps just use arrows and words on the figures for the dots on those two, and eliminate them from the legend?  

Accept and made suggested changes (Figure 1)

Figure 2 – this makes sense.

Section 3.2.1:  Either here or above where I noted, it would be good to indicate the spatial footprint of the phenocam, to get some sense of the representativeness of the data relative to pixels.

Accept and added above. Thanks.

3.2.2  Seems the title of this section should be different than the prior

Good catch. We changed to title of this section to eMODIS NDVI and phenology metrics. We also corrected 3.1

Given the argument for the paper – that we need to see if moderate resolution data can get phenology -- it’s confusing to me that we would be testing it against a dataset that is somehow influenced by MODIS data.   More justification in this section would be helpful.

We felt this testing was necessary to understand how the ‘standard’ phenology products compared to the higher resolution HLS. In the discussion and conclusions, we noted the shortfalls of the HLS due to the temporal frequency and cloudiness of the region. We do not feel we could reach these conclusions regarding data frequency without the context provided by the weekly eMODIS phenological results.

3.2.3  This section is confusingly written.  The expectation I had leading up to this section and 3.2.2 was that we’d be comparing the HLS results directly to the phenocam;  thus, the introduction of 250-m based analysis is odd.  I think I can infer what’s going on, but that’s partly guessing.  Also, the specifics of the workflow in this section are unclear.  So, how is the comparison actually happening?   We need that logic first, but it’s not really stated anywhere yet. Then we can be informed about the need to only do that at pixels dominated by grassland, and finally how you built the mask for that purpose.

We added more detail to the section 3.2.2 to clarity the concept of using eMODIS. We also added more description to the phenology method section to illustrate the workflow (line 257 - 262). We hope these changes will resolve this clarity issue.

3.4:   It is not clear to me how precipitation 34km away from the wetland site will be used for reference data – more explanation here would be helpful.  This is particularly necessary because the argument of the paper for moderate resolution information would seem – at first glance, at least – to be negated by comparing against distant data.  The whole wetland area would necessarily be treated the same, right?  In that case, couldn’t we just use a coarser resolution sensor?  I suspect not, but more logic to bolster your argument would help.

The wetlands in this region vary in size but many are quite small. Moderate resolution may capture less than half (this is an estimate), thus the benefit of using Landsat or Sentinel scale data. You point regarding precipitation is correct--most of the wetlands would be subject to similar precipitation using the data we included. Runoff and ground water tend to have more influence on wetland differences while the overall pattern during a season is influenced by moisture availability.  To address some of your concern, we incorporated the monthly precipitation data from NOAA for the county where the study area is located (line 237 – 239, Figure 11). We also added the description of the water surface elevation data (line 240 - 245).

Line 240:  suggest “temporally normalized using a compositing strategy described in the steps below:”

Accept and changed (line 257)

Line 245:  It takes a leap of imagination to catch the fact that this is the compositing strategy that results in making the time step the same – clarify this a bit more. 

We modified the description as “The composition strategy selected maximum NDVI values from either Landsat-8 or Sentinel-2 for each pixel for each week across the entire time period.” (line 265 - 266) We hope these changes will resolve this clarity issue.

Step 7:  this seems to describe the comparison with the eMODIS data, but not the phenocam.  Was just a single 30m pixel used to compare with Phenocam? 

Yes. A single 30-m Landsat pixel over the PhenoCam site was used to compare with Phenocam data. A sentence was added to step 7 to make this point clear (line 286 - 288).

272:  Add in which component of CCDC’s outputs you actually considered to be the results – presumably the disturbance date, but please specify.

We added “the change dates from the CCDC results” to the text for clarity (line 295 - 296).

Figure 3:  I appreciate the authors being honest in showing these results, but they certainly do not cast the moderate resolution data in a very favorable light.  At first glance, it appears the eMODIS data are capturing variability in spatial patterns commensurate with the likely spatial patterns of the fields, and that the temporal frequency drives all of the pattern in the moderate resolution products. 

You are correct that the temporal frequency does influence the results and we included discussion regarding this issue. Although at the course scale, eMODIS data appear to capture spatial pattern, the subpanels (e, and f) suggest that spatial patterns at the local scale are better captured by the 30 m data in the high data frequency zone while the eMODIS is more generalized (line 345 - 347).

333:  This is confusingly written, I think – do you mean that you took the 250-m grassland cells, identified all of the 60+ Landsat pixels within that and calculated the stdev of SOST across those Landsat pixels? Figure 5 would suggest that, but I’m not clear. 

Yes. The procedure was described in described in step 7 and we added additional text for clarity as mentioned in a prior response. We don’t feel that incorporating more text here is now necessary.

Figure 4:  I appreciate the argument about there being a tighter distribution.  However, the motivation for the paper appears to me to be very spatial-scale-specific – can moderate resolution data do better at getting spatial patterns without the problems of mixed pixels that occurs with MODIS-scale pixels?   To answer that – especially given the patterns in figure 3 – we would want to see a scatter plot of MODIS SOST vs mean(median) SOST from corresponding cells in HLS data, not just the overall regional-scale means and distributions.  Just limiting it to the grassland pixels is fine, since the map in Figure 3 presumably contains a bunch of non-grassland pixels that confuse things.   Maybe the grassland cells do compare well on a pixel-by-pixel basis?  Upshot: I don’t think that these boxplots are the strongest arguments for this approach.   They can be kept in, but pixel-to-pixel comparisons seem more appropriate for the argument of the paper.

We appreciate your suggestion. The scatter plot is added as Figure 4 with additional descriptions (line 354 - 361).

348 : typo on Vierling’s name

We confirmed with Google Scholar that the author is Anton Vrieling and the spelling for the reference is correct.

356:  nitpicky, I realize, but the HLS time series timingcorresponded well, but the NDVI absolute values were quite different.  Since timing is what you’re after, this is fine, but it would help to be precise in the language here since this is a central finding.

If we understand your comment, you suggest that the HLS data seem to capture the phenological patterns. We agree with your observation for instances where the data was dense enough. The magnitude of NDVI varied between sensors. NDVI values are lower from the phenocam because of BDRF (the camera points north so Sun is illuminating the canopy from “behind” the phenocam). We suspect the differences between the Landsat/HLS and eMODIS magnitude is related to resolution differences. This is an interesting finding but one we felt would be best explored in a future paper.

Figure 6—please make the font size of the legend larger – there is room on the graph, and right now it’s quite small; same for the x-axis.

Accepted. We increased the font size of the legend, x-axis, and y-axis (Now Figure 7).

The finding of the figure is encouraging, though!!! 

Thanks!

Figure 7 – please put scale bars on these.  Also, for subpart c) indicate that the reference products are MTBS and BAECV – the figure should stand alone without reference back to the main text. 

Accepted. Scale bars added, and the figure description now includes MTBS and BAECV (Now Figure 8).

It would be instructive to see some actual pixel-level trajectories to see why the early season fires were detectable using individual Landsat images for MTBS and BAECV, but not the time series.  

Accepted suggestion and added as Figure 9 with a figure description included (line 412 - 415).

390:  water surface elevation data?   These are not mentioned earlier, but they seem to be a critical reference dataset that would alleviate my concerns about only validating relative to a distant weather station.

Thank you for pointing out this oversight. We added a description of the water elevation data in the source data section of the paper (3.4.2).

Figure 8 – this is a useful figure!

Thanks!

423-431 – this is a good summary of what was found.

Thanks.

519-521—Understood, but it would be useful to comment on whether it makes sense to use these data vs MODIS data, since that is how the introduction seems to set up our expectations.  And to be clear, you do state at the beginning of the conclusion section that you’re testing whether the Landsat-8 plus Sentinel provides “sufficient frequency … to enable monitoring seasonal dynamics…” – that suggests that you’re not just testing whether there is improvement, but whether the underlying processes can be tracked.

We feel that our findings could help guide whether it makes sense to use HLS data or eMODIS data in future research. Making a specific recommendation would be difficult as the data choice is dependent on the site characteristics (homogenous verses heterogeneous) and HLS data availability and frequency of observations. We expect this may pose challenges in the near-future due to data limitations and the current acquisitioning schedule. These points were covered within the discussion and conclusions.

Reviewer 2 Report

 The study summarized in the manuscript would be a great interest to many researchers. The remotely sensed data fusion technique the authors explored has a great potential to mitigate known challenges, data gap, in remotely sensed time series data analysis. Outlining methods by the step would provide audience a clear instructions to apply the method discussed.  Interpretation of results are clearly stated and applicability of the method under assumption for future satellite missions is very insightful.  I wonder what next step or future studies the authors plan to advance current study.  The only concern I have is the font size in many figures is too small to read.  Also there are missing labels for legends and axis for some figures.

Author Response

We would like to thank the reviewers for their constructive comments and suggestions. We have revised the manuscript accordingly. Please find attached a point-by-point response to reviewer’s concerns. We hope that you find our responses satisfactory and that the manuscript is now acceptable for publication.

Reviewer 2:

The study summarized in the manuscript would be a great interest to many researchers. The remotely sensed data fusion technique the authors explored has a great potential to mitigate known challenges, data gap, in remotely sensed time series data analysis. Outlining methods by the step would provide audience a clear instructions to apply the method discussed.  Interpretation of results are clearly stated and applicability of the method under assumption for future satellite missions is very insightful.  I wonder what next step or future studies the authors plan to advance current study.  The only concern I have is the font size in many figures is too small to read.  Also there are missing labels for legends and axis for some figures.

Accepted.

We reproduced Figure3 and 6 with increased font size.

We also reproduced Figure1, 7, and the supplement figure A2 to make the text clearer.

We added a final paragraph to suggest future following research.