Shifted Global Vegetation Phenology in Response to Climate Changes and Its Feedback on Vegetation Carbon Uptake

Round 1

Reviewer 1 Report (Previous Reviewer 4)

I appreciate that the authors took the time to answer my questions, however, they did NOT address any of my comments or concerns on the uncertainties in the remote sensing and Fluxnet GPP data. The steps the authors took were: 1) quality control to remove outliers, 2) curve fitting and interpolating the 8-day data to daily, and 3) calculating dGPP/dt (FOD) using the interpolated daily data. They then defined SOS and EOS as maximum and minimum FOD except for ecosystems showing multiple growing seasons. However, there are a number of uncertainties in that process: 1) MODIS data is an 8-day composite which means the data recorded is the “best value” that could have occurred in any measurement within the 8-day period, 2) MODIS GPP is based on reflectance measurements and assumptions about land cover types and ecosystem structure which means it suffers from the same uncertainties presented in the vegetation index measurements plus the assumptions themselves, and 3) both removing the outliers and interpolation could have influenced the shape of the fitted daily GPP curve, and therefore change the derived FOD and the FOD-based SOS and EOS. Each of these can cause a bias of several days which are at a similar scale to the presented RMSE and ME and much larger than the derived trends of 0.19 days/year, and yet, the authors did NOT address any of the above-listed uncertainties.

That also applies to the Fluxnet GPP estimates. I check Peng et al (2017), and I am not sure they said anywhere that “the different GPP extracts were identical in terms of phenology and did not make a difference”. In fact, they presented considerable differences between SOS based on MODIS-derived VIs (NDVI and EVI) and SOS from ground observations (NPN) and AmeriFlux. Moreover, a more relevant reference on Fluxnet GPP suggested at least using both daytime (DT) and nighttime (NT) variables to assess uncertainty.

In the response, the authors claimed that “we think for most parts of the globe, the 8-days GPP data can well reflect the critical events in vegetation growth cycle” and “we believe that the comparison between the modeled and observed result provided the basic information for the uncertainty”. However, that does not answer my questions or address my concerns and I did not see any evidence in either the response or the revised manuscript why they believed that or more importantly, why the readers should believe them.

 

A few additional comments on the revised manuscript:

Fig 2 and Response 8 (d) “We understand for a particular point (location), this may involve uncertainty, but we also believe that results from a large number of locations will largely minimize the uncertainty and can provide useful patterns”: I do not know which point this is, but based on my experience working with MODIS data, that is actually pretty common. Again, what you believe is not enough to convince readers.   

Fig 3: are the data points site-year or site average?

Lines 584-586 “Additionally, a flexible range of 10% relative difference was given to pick up the first and last FOD points if the values are within that range for MGS cases”:  I am not entirely sure this means the same thing as you said before in lines 261-264: “For ecosystems showing multiple growing seasons within an annual growth cycle, the critical FOD values are taken to be the first and last FOD points with values less than 10% of the maximum and minimum in relative difference within that cycle”. That is, is the difference 10% or 90%? And could you show what that looks like in a figure?

In addition, define “much lower” in lines 259-260: “The critical FOD values take the maximum and minimum FOD if all the rest FOD points in an annual vegetation growth 260 cycle present much lower values”.

References:

 

Pastorello, G., Trotta, C., Canfora, E., Chu, H., Christianson, D., Cheah, Y. W., ... & Law, B. (2020). The FLUXNET2015 dataset and the ONEFlux processing pipeline for eddy covariance data. Scientific data, 7(1), 1-27.

Author Response

I appreciate that the authors took the time to answer my questions, however, they did NOT address any of my comments or concerns on the uncertainties in the remote sensing and Fluxnet GPP data. The steps the authors took were: 1) quality control to remove outliers, 2) curve fitting and interpolating the 8-day data to daily, and 3) calculating dGPP/dt (FOD) using the interpolated daily data. They then defined SOS and EOS as maximum and minimum FOD except for ecosystems showing multiple growing seasons. However, there are a number of uncertainties in that process: 1) MODIS data is an 8-day composite which means the data recorded is the “best value” that could have occurred in any measurement within the 8-day period, 2) MODIS GPP is based on reflectance measurements and assumptions about land cover types and ecosystem structure which means it suffers from the same uncertainties presented in the vegetation index measurements plus the assumptions themselves, and 3) both removing the outliers and interpolation could have influenced the shape of the fitted daily GPP curve, and therefore change the derived FOD and the FOD-based SOS and EOS. Each of these can cause a bias of several days which are at a similar scale to the presented RMSE and ME and much larger than the derived trends of 0.19 days/year, and yet, the authors did NOT address any of the above-listed uncertainties.

Response: The authors appreciate your comments and understand the concerns. At the global scale, the 8-days GPP data is the only feasible way, in terms of the spatial coverage and temporal resolution, to map vegetation dynamics and vegetation growth cycle. We had no alternative remote sensing products that can provide a comparable result, in terms of the spatial coverage and temporal resolution, to realize mapping vegetation dynamics globally. There have been numerous studies applying MODIS GPP data in mapping vegetation phenology, though extensive work has yet to be done at global scale and with focus on their responses to climate changes and feedback on vegetation productivity.

We thank you for pointing out the sources of the uncertainties. It is inevitable that a specific year there could be a considerable bias in the phenological dates. However, from statistical point of view and to the best knowledge of the authors, if we combine time-series data to draw a long-term trend (here n=20), the result can still be meaningful, particularly if that trend has passed statistical significance test.

For the three specific points inducing uncertainties as mentioned, we have the following answers.

1) MODIS data is an 8-day composite which means the data recorded is the “best value” that could have occurred in any measurement within the 8-day period.

Let’s suppose that the date is randomly distributed in one of the 1~8 days in each period, considering that the annual curve was fit from 40+ points (periods) in each year, the uncertainty due to the random effect in each period will be largely minimized. Applying 8-days MODIS GPP to derive vegetation is actually widely used in literature, though we agree this product is not perfect and further improvement is needed.

2) MODIS GPP is based on reflectance measurements and assumptions about land cover types and ecosystem structure which means it suffers from the same uncertainties presented in the vegetation index measurements plus the assumptions themselves.

We appreciate the comment. In fact, there have been disputes over the GPP or vegetation index measurements (e.g., NDVI) in mapping vegetation phenology. GPP is the most direct indicator to vegetation phenology as it represents the growth rate of vegetation, but we agree that model-based GPP may contain more uncertainties due to modelling errors while NDVI is obtained directly from remote sensing data without suffering extra intermediate steps which can introduce bias. NDVI is a useful indicator that reflect the greenness of vegetation and is closely related to vegetation growth rate. In a word, GPP is the direct indicator to vegetaion phenology and NDVI is an indirect indicator. From this perspective, GPP is more preferred choice. Recent study indicates that phenological characteristics derived from Solar-Induced Chlorophyll Fluorescence (SIF, an indicator proxy of GPP) than VI (Wang et al., 2022. Comparison of Phenology Estimated From Monthly Vegetation Indices and Solar-Induced Chlorophyll Fluorescence in China, Front. Earth Sci.). The authors did a preliminary literature review and found that existing studies even favoured GPP more instead of NDVI.

3) both removing the outliers and interpolation could have influenced the shape of the fitted daily GPP curve, and therefore change the derived FOD and the FOD-based SOS and EOS.

Outliers are errors and should be rectified in reasonable ways. We followed widely recognized approach to process the outliers. Removing outliers and interpolation are actually prerequisite steps for mapping vegetation phenological events from remote sensing imagery. Those processes were to minimize uncertainties in the remote sensing dataset. Almost all studies applying remote sensing imagery need to take those steps. We tested numerous times and given choices either removing the outliers (and interpolation) or not, we definitely prefer to taking such steps; otherwise the results could suffer more uncertainties from the outliers.

2. That also applies to the Fluxnet GPP estimates. I check Peng et al (2017), and I am not sure they said anywhere that “the different GPP extracts were identical in terms of phenology and did not make a difference”. In fact, they presented considerable differences between SOS based on MODIS-derived VIs (NDVI and EVI) and SOS from ground observations (NPN) and AmeriFlux. Moreover, a more relevant reference on Fluxnet GPP suggested at least using both daytime (DT) and nighttime (NT) variables to assess uncertainty.

Response: We are sorry that the reference was actually updated due to the revision and the sequence was wrongly mentioned in the last response letter (it was originally listed as [6] but updated to [8] after we made the revision). The correct reference is: Wang, X.; et al, No Trends in Spring and Autumn Phenology during the Global Warming Hiatus. Nat. Commun. 2019, 10, 1–10, doi:10.1038/s41467-019-10235-8.

In this paper it mentioned that “the different GPP extracts were identical in terms of phenology and did not make a difference” in the last sentence in the Methods section (Phenology estimation from the FLUXNET database)

In addition, in appendix Fig. 20, the paper clearly showed that daytime (DT) and nighttime (NT) obtained the same result. Please see the figure below (Please refer to the attachment because the figure cannot be inserted here):

3. In the response, the authors claimed that “we think for most parts of the globe, the 8-days GPP data can well reflect the critical events in vegetation growth cycle” and “we believe that the comparison between the modeled and observed result provided the basic information for the uncertainty”. However, that does not answer my questions or address my concerns and I did not see any evidence in either the response or the revised manuscript why they believed that or more importantly, why the readers should believe them.

Response: We are sorry we did not paraphrase the terms well. For “we think for most parts of the globe, the 8-days GPP data can well reflect the critical events in vegetation growth cycle”, we would like to emphasize that the 8-days GPP data can well reflect the critical events in vegetation growth cycle. The reason is that at least 70% of existing studies mapping vegetation growth and growth cycle at regional scales or large scales utilized the 8-days GPP data, though the steps and models in each work may differ. Moreover, for “we believe that the comparison between the modeled and observed result provided the basic information for the uncertainty”, we think that the comparison between modelled and observed result provides essential and objective information for assessing the uncertainty. We may resort to simulation or other mathematical models to assess the uncertainty but we believe that observation is the most reliable method. We think our statements are based on existing literature and common sense and that readers understand what we mean.

4. A few additional comments on the revised manuscript:

Response: Fig 2 and Response 8 (d) “We understand for a particular point (location), this may involve uncertainty, but we also believe that results from a large number of locations will largely minimize the uncertainty and can provide useful patterns”: I do not know which point this is, but based on my experience working with MODIS data, that is actually pretty common. Again, what you believe is not enough to convince readers.   

We are sorry to cause misunderstanding. The “point” here means spatial location. What we meant was that we might find varied uncertainty for a modelled target variable (for example, GPP or vegetation phenological dates) at a single point (pixel) from remote sensing imagery, but when we have multiple time-series data over a large area, the pattern from the data will be clear and useful.   

Fig 3: are the data points site-year or site average?

They are site-year basis. We have provided explanation in the title of the figure.

Lines 584-586 “Additionally, a flexible range of 10% relative difference was given to pick up the first and last FOD points if the values are within that range for MGS cases”:  I am not entirely sure this means the same thing as you said before in lines 261-264: “For ecosystems showing multiple growing seasons within an annual growth cycle, the critical FOD values are taken to be the first and last FOD points with values less than 10% of the maximum and minimum in relative difference within that cycle”. That is, is the difference 10% or 90%? And could you show what that looks like in a figure?

They refer to the same thing. The following shows an example (please refer to the attachment as the figure cannot be inserted here):

We compute the maximum and minimum for FOD at all critical points first. For this case, the maximum and minimum FOD would be Max=MaxFOD(F1,F2,F3,F4) and Min=MinFOD(F1,F2,F3,F4), respectively. If FOD@F1, or FOD at point F1, satisfies |Max-FOD@F1|<10%* Max, then F1 will be selected as the starting point (SOS), even FOD@F1<FOD@F3; otherwise the next point F3 will be selected for SOS (there are only two potential points, F1 and F3, for SOS but the principle applies to points>2). Similarly, the last points satisfying |Min-FOD@Fx|<10%*|Min| will be selected as EOS.  Intuitively, Max=MaxFOD(F1,F2,F3,F4)=FOD@F3. If the FOD@F1 is very close to Max (less than 10% in relative difference), the first point will be set as SOS, otherwise the point with the largest FOD is set as SOS. The same principle applies to EOS.

In addition, define “much lower” in lines 259-260: “The critical FOD values take the maximum and minimum FOD if all the rest FOD points in an annual vegetation growth 260 cycle present much lower values”.

This section refers to the same thing. Take again the above figure as example. Since F3 presents the maximum value for all the potential points for SOS, if all the other points (here only F1) have FOD values less than 90% (1-10%) of that at F3 in relative difference, only F3 can be taken as SOS. Similar principle applies to EOS.

We have updated the text in this revised version.

We again thank you for your valuable comments and we have tried best to improve the paper. We hope that our revision could address your concerns.

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

The submitted manuscript documents a global analysis of Gross Primary Productivity (GPP)-based Land Surface Phenology (LSP) during 2001-2020. Utilizing the MODIS GPP data product (MCD17A2H), the authors present a phenology algorithm based on filtering, smoothing, fitting, and finding the first-order derivative. The authors compare these phenology metrics to those obtained from a similar algorithm applied to flux tower data. Finally, they evaluate changes in global phenology over the time period studied, the contributions of temperature and precipitation to the LSP changes, and the alignment between the length of season and the annual GPP, both from the MODIS GPP product.

 

The manuscript has several strengths. The research area is incredibly important to current questions in climate and Earth science. Biogeophysical and biogeochemical feedbacks between climate and vegetation, such as changes in annual GPP due to shifts in LSP, are key sources of uncertainty in climate projections. The data used in this study are temporally and spatially extensive and consistent. The method of identifying trends in phenology using Theil-Sen slope and Mann-Kendall significance is appropriate.

 

The manuscript also has some weaknesses. The authors highlight results that are not filtered for significance and thus have little meaning. Although the authors calculate the Mann-Kendall significance of phenology trends, and do present results filtered for significance, these statistics are buried after the unfiltered results and not mentioned in the abstract (while the unfiltered results are).

 

I also have concern about the partial correlation analysis between the GPP-based LSP metrics and monthly average temperature and total precipitation from ERA5-Land. The MODIS algorithm for calculating GPP relies upon meteorological data from reanalysis (GMAO/NASA) in its predictions. These meteorological inputs are used to scale the radiation use efficiency and they are particularly impactful on the GPP predictions around the start and end of the growing seasons. Therefore, to identify correlation between GPP-based LSP metrics and meteorological data may not be meaningful in the way the authors expect. Rather, it seems that they are simply detecting the effects of meteorological data used in the algorithm.

 

Suggested revisions:

·      Because the LSP algorithm used in this paper is based upon GPP, the authors should clarify throughout the manuscript which time-scale of GPP they are referring to. For example, when comparing the LOS to the GPP, they should clarify that they are comparing to the annual GPP.

·      The authors should place much more emphasis on the statistical significance of their findings. Findings without statistical significance should be removed entirely or moved to the appendix. Only trends with statistical significance should be shared in the abstract and conclusions. The authors should also read carefully for how they use the word “significant” and remove any instances that do not strictly mean statistical significance at p < 0.5, such as the instance in the abstract. Another particularly concerning use is in the caption of figure 10.

·      The validation of the GPP-based LSP algorithm used with the ground truth from FLUXNET should be strengthened (shown in figure 3). The correlations shown appear weak for drawing sweeping conclusions about global shifts in phenology or climate-vegetation feedbacks. I suggest that the authors review the comparison between remotely sensed LSP and GPP-based phenology metrics from eddy covariance-based shown in Gao (2023). This comparison separates by biome, indicating which vegetation types the algorithm performs strongly in, and which it does not.

·      I suggest that the authors review the ATBD for the MODIS GPP algorithm and reconsider their approach of examining correlations among the MODIS GPP-based LSP metrics and temperature and precipitation. Can they instead find an independent phenology metric to compare with the meteorology?

 

 

Gao, Xiaojie, Ian R. McGregor, Josh M. Gray, Mark A. Friedl, and Minkyu Moon. "Observations of satellite land surface phenology indicate that maximum leaf greenness is more associated with global vegetation productivity than growing season length." Global Biogeochemical Cycles (2023): e2022GB007462.

Author Response

The submitted manuscript documents a global analysis of Gross Primary Productivity (GPP)-based Land Surface Phenology (LSP) during 2001-2020. Utilizing the MODIS GPP data product (MCD17A2H), the authors present a phenology algorithm based on filtering, smoothing, fitting, and finding the first-order derivative. The authors compare these phenology metrics to those obtained from a similar algorithm applied to flux tower data. Finally, they evaluate changes in global phenology over the time period studied, the contributions of temperature and precipitation to the LSP changes, and the alignment between the length of season and the annual GPP, both from the MODIS GPP product.

The manuscript has several strengths. The research area is incredibly important to current questions in climate and Earth science. Biogeophysical and biogeochemical feedbacks between climate and vegetation, such as changes in annual GPP due to shifts in LSP, are key sources of uncertainty in climate projections. The data used in this study are temporally and spatially extensive and consistent. The method of identifying trends in phenology using Theil-Sen slope and Mann-Kendall significance is appropriate.

Response: We very much appreciate your comments.

 

The manuscript also has some weaknesses. The authors highlight results that are not filtered for significance and thus have little meaning. Although the authors calculate the Mann-Kendall significance of phenology trends, and do present results filtered for significance, these statistics are buried after the unfiltered results and not mentioned in the abstract (while the unfiltered results are).

Response: Thank you for your comments and we have addressed the concerns in this revised paper. Please refer to the following specific points.

 

I also have concern about the partial correlation analysis between the GPP-based LSP metrics and monthly average temperature and total precipitation from ERA5-Land. The MODIS algorithm for calculating GPP relies upon meteorological data from reanalysis (GMAO/NASA) in its predictions. These meteorological inputs are used to scale the radiation use efficiency and they are particularly impactful on the GPP predictions around the start and end of the growing seasons. Therefore, to identify correlation between GPP-based LSP metrics and meteorological data may not be meaningful in the way the authors expect. Rather, it seems that they are simply detecting the effects of meteorological data used in the algorithm.

Response: Thank you for this comment. Modeling MODIS GPP relies upon meteorological variables but they are not linearly related to GPP in the model. In other words, even meteorological variables are components in the input for computing MODIS GPP, they are not supposed to be linearly related to the GPP. If we simply analyze the meteorological variables that are input into the GPP model and output of GPP from the model, we cannot, mathmatically and practically, draw any a linear relationsihp between the output and the input. In fact, in majority of the study area we found a nonlinear relationship between the variables in the correlation alaysis. Threfore, because the correlation analysis reflects a linear relationship between the two variables, there is no direct link between the model and the output from the correlation analysis. The linear relationship, if found, actually provides a different mechanism between the variables but does not come from the input-output link in the model.

Suggested revisions:

Because the LSP algorithm used in this paper is based upon GPP, the authors should clarify throughout the manuscript which time-scale of GPP they are referring to. For example, when comparing the LOS to the GPP, they should clarify that they are comparing to the annual GPP.

Response: Thank you for pointing out this problem in manuscript. We have clarified the time scale of GPP in the manuscript. When comparing the LOS to the GPP in section 3.4, we specify that this is the annual GPP.

The authors should place much more emphasis on the statistical significance of their findings. Findings without statistical significance should be removed entirely or moved to the appendix. Only trends with statistical significance should be shared in the abstract and conclusions. The authors should also read carefully for how they use the word “significant” and remove any instances that do not strictly mean statistical significance at p < 0.5, such as the instance in the abstract. Another particularly concerning use is in the caption of figure 10.

Response: We gratefully appreciate for your valuable suggestion. We corrected the statements of significance in the abstract and conclusions, and we rephrased the results by keeping the figure from the statistical significance. We also corrected the caption of Figure 10.

The validation of the GPP-based LSP algorithm used with the ground truth from FLUXNET should be strengthened (shown in figure 3). The correlations shown appear weak for drawing sweeping conclusions about global shifts in phenology or climate-vegetation feedbacks. I suggest that the authors review the comparison between remotely sensed LSP and GPP-based phenology metrics from eddy covariance-based shown in Gao (2023). This comparison separates by biome, indicating which vegetation types the algorithm performs strongly in, and which it does not.

Response: Thank you for the above suggestions. The validation results of the separate vegetation types are reflected in the RMSE and ME (Figure 3 (a), (b)), which shows how the algorithm performs in different vegetation.

I suggest that the authors review the ATBD for the MODIS GPP algorithm and reconsider their approach of examining correlations among the MODIS GPP-based LSP metrics and temperature and precipitation. Can they instead find an independent phenology metric to compare with the meteorology?

Response: We appreciate the comment. In fact, there have been disputes over the GPP or vegetation index measurements (e.g., NDVI) in mapping vegetation phenology. GPP is the most direct indicator to vegetation phenology as it represents the growth rate of vegetation, but we agree that model-based GPP may contain more uncertainties due to modelling errors while NDVI is obtained directly from remote sensing data without suffering extra intermediate steps which can introduce bias. NDVI is a useful indicator that reflect the greenness of vegetation and is closely related to vegetation growth rate. From existing studyies, GPP is the direct indicator to vegetaion phenology and NDVI is an indirect indicator. From this perspective, GPP is more preferred choice.However, GPP could suffer from more uncerties. That would place a balance for the selection of the variable. Recent study indicates that phenological characteristics derived from Solar-Induced Chlorophyll Fluorescence (SIF, an indicator proxy of GPP) than VI (Wang et al., 2022. Comparison of Phenology Estimated From Monthly Vegetation Indices and Solar-Induced Chlorophyll Fluorescence in China, Front. Earth Sci.). The authors did a preliminary literature review and found that existing studies even favoured GPP more instead of NDVI.

Author Response File: Author Response.docx

Reviewer 3 Report (New Reviewer)

This manuscript attempts to analyze mainly the spatial and temporal changes in global vegetation phenology, examining the response of vegetation phenology to climate change and the feedback effects of vegetation phenology changes on GPP. The topic is appealing to the reader, but the manuscript can be improved.

--- The authors derived the phenological metrics of SOS, EOS and LOS from GPP data. They then analyzed the relationships between the phenological metrics and GPP in Figure 8. What is the relative contribution of the their common components (i.e., GPP) on the results?  How to eliminate it? They should tried to improve this issue.

--- The color of FLUX sites in Fig. 1 is not obvious enough, and it is suggested to use a color with more contrast with the background.       

Author Response

This manuscript attempts to analyze mainly the spatial and temporal changes in global vegetation phenology, examining the response of vegetation phenology to climate change and the feedback effects of vegetation phenology changes on GPP. The topic is appealing to the reader, but the manuscript can be improved.

Response: We appreciate the comments and we improved the manuscript as documented in the following and many others.

 

Point 1: The authors derived the phenological metrics of SOS, EOS and LOS from GPP data. They then analyzed the relationships between the phenological metrics and GPP in Figure 8. What is the relative contribution of the their common components (i.e., GPP) on the results?  How to eliminate it? They should tried to improve this issue.

 

Response: The phenometrics were modeled from the time series dynamics (the shape) of an annual GPP cycle and based on threshold method. Conversely the correlation or the influence on GPP from vegetation phenology was linearly modeled depending on the average or accumulated values for both variables. In other words, the practice of extracting the phenometrics from GPP dynamics does not suggest they will have linear relationship between the two variables. In fact, as proved from our study, majority of the study area did not show significant correlation between the two variables.

In this revised version, we have supplemented more information of the possible uncertainties related to the results.

 

Point 2: The color of FLUX sites in Fig. 1 is not obvious enough, and it is suggested to use a color with more contrast with the background.

Response: As Reviewer suggested, we have set the color of the FLUX sites to a more contrasting color to the background.

Author Response File: Author Response.docx

Reviewer 4 Report (New Reviewer)

The authors used MOD17A2H GPP product to extract the global vegetation phenology, and analyze their spatio-temporal patterns among different vegetation biomes. The methods and results are given in detail in this paper, and the main suggestions are as follows.

(1) Why did you use GPP in place of NDVI to extract vegetation phenology? What is the advantage of GPP for phenolgy extraction? Please add the difference beween GPP-based phenology and NDVI-based phenology, as well as their ecological significance in the Introduction section.

(2) As we know, MODIS GPP  is not an observation variable like remote sensing NDVI or flux-based GPP,  but a product from ecological model combing climate variables and remote sensing datasat.  Please analyze the uncertainties from simulated GPP.

(3) Is the method in this study suitable for the multiple cropping in one  year? Please further analyze  the uncertainties of cropland and try to modify the phenology extraction of some typical crops.

•  

•  

•  

Author Response

Why did you use GPP in place of NDVI to extract vegetation phenology? What is the advantage of GPP for phenolgy extraction? Please add the difference beween GPP-based phenology and NDVI-based phenology, as well as their ecological significance in the Introduction section.

Response: Thank you so much for the comment. While traditional vegetation index-based phenology reflects changes in vegetation greenness, vegetation phenology extracted from GPP time series (also known as vegetation photosynthetic phenology) can more accurately capture changes in vegetation photosynthetic activity, which is more important for studying carbon sequestration in vegetation and terrestrial ecosystems. Fang, et al, 2022 (Vegetation photosynthetic phenology metrics in northern terrestrial ecosystems: a dataset derived from a gross primary productivity product based on solar-induced chlorophyll fluorescence. Earth System Science Data Discussions, 1-39, 2022) mentioned in their study that vegetation index-based phenology is not sufficient to represent the seasonal characteristics of photosynthesis, so we chose GPP instead of NDVI to extract vegetation phenology. We have added these descriptions in the introduction section.

 

As we know, MODIS GPP is not an observation variable like remote sensing NDVI or flux-based GPP, but a product from ecological model combing climate variables and remote sensing datasat. Please analyze the uncertainties from simulated GPP.

Response: Fang has compared the differences between GPP-based phenology and vegetation index-based phenology in their study. The correlation between the GPP-based phenology and the phenology based on carbon fluxes was better and the RMSE was smaller compared to the phenology obtained from NDVI/EVI. And more detailed information can be found in their article (such as Table 1 of the mentioned article).

In this revised version, we have provided the uncertainty information in Fig. 5, Fig. 10, and the discussion section 4.2.

Is the method in this study suitable for the multiple cropping in one year? Please further analyze the uncertainties of cropland and try to modify the phenology extraction of some typical crops.

Response: Thank you for pointing out this common issue in vegetation phenology mapping from remote sensing imagery. Extracting phenology for multiple growing seasons (MGS) in Cropland is a challenge. We used threshold value, 10% from the MaxFOD, to decide if vegetation growth dynamics is MGS or not.

Given the above example (see attached file because figures cannot be inserted) which could be potentially decided as MGS, the maximum and minimum FOD would be Max=MaxFOD(F1,F2,F3,F4) and Min=MinFOD(F1,F2,F3,F4), respectively. If FOD@F1, or FOD at point F1, satisfies |Max-FOD@F1|<10%* Max, then F1 will be selected as the starting point (SOS), even FOD@F1<FOD@F3; otherwise the next point F3 will be selected for SOS (there are only two potential points, F1 and F3, for SOS but the principle applies to points>2). Similarly, the last points satisfying |Min-FOD@Fx|<10%*|Min| will be selected as EOS.  Intuitively, Max=MaxFOD(F1,F2,F3,F4)=FOD@F3. If the FOD@F1 is very close to Max (less than 10% in relative difference), the first point will be set as SOS, otherwise the point with the largest FOD is set as SOS. The same principle applies to EOS.

We documented the method in section 2.2.3 and provided further discussion in section 4.2.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report (New Reviewer)

I appreciate that the authors have responded to all of my concerns over the previous version of the manuscript. In particular, their edits to add clarification of the timescale of the GPP used in comparison with the phenology metrics was quite helpful. However, my primary concerns about the manuscript remain. In brief, these are (1) results not filtered for significance are presented in the abstract as major findings, and (2) the comparison of phenology metrics based on the MODIS GPP product with meteorological data cannot identify previously-unknown relationships between phenology and climate. I elaborate these remaining concerns below.

 

The authors indeed edited the abstract to highlight only the significant results from the comparison of annual GPP with phenology. However, they continue to highlight other results in the abstract that are not filtered for significance. Here is the sentence that concerns me: “Results indicated that SOS advanced significantly, at an average rate of 0.19 days/year at global scale, particularly in the northern hemisphere above the middle latitude (≥30°N) and that EOS was slightly delayed during the past two decades, resulting in prolonged LOS in 72.5% of the vegetated area.” This sentence should be rewritten to present only those changes in SOS, EOS, and LOS which were identified as significant using the Mann-Kendall test described in the manuscript. Specifically, the authors should remove reference to an average rate of change in SOS which was not verified using the Mann-Kendall test and remove reference to the percent of area seeing changes in LOS which were not verified with the Mann-Kendall test.

 

My primary remaining concern is with the authors’ approach to addressing the second of their three objectives: “Does the climate factors affect vegetation phenology among different vegetation types? The responses of vegetation phenology to climate fluctuations can be examined based on the correlation analysis on the spatial variables.” The authors claim in their response to my original concern that the purpose of their correlation analysis is specifically to identify the linear nature of known and accepted relationships between climate and phenology. This is rather different from the described goal of this step, which nowhere mentions their particular focus on linearity.

 

This distinction is important because a large body of work exists on the responses of land surface phenology to changes in climate (Liu, et al. 2016, Chen, Yang, and Du 2022, Gao and Zhao 2022, Li, et al. 2022, Shen, et al. 2022, Zhang, et al. 2022, Zimmer, et al. 2022).  To add to this nuanced understanding, studies should include novel data or methods, or focus deeply on a specific site or context. If the authors are attempting to make a contribution by studying the specifically linear nature of some known relationships, then they should make that clear in the presentation of the study, when stating the objective, describing the methods, and sharing the results.

 

As currently written, the authors give the impression that their analysis will be able to identify previously-unknown relationships between climate and phenology. I would argue, and it sounds like the authors would agree, that this is not possible when using the MODIS-based GPP data set because it is based upon the assumption of known relationships between meteorological data and GPP. While the MODIS GPP algorithm may produce GPP phenology with either linear or non-linear relationships to climate, it cannot produce GPP phenology with relationships to climate that are independent of its inputs (meteorology, greenness, and land cover/vegetation) and algorithm.

 

If the authors feel that their results are an important contribution because they identify specifically linear relationships between phenology based on MODIS GPP and climate, then I recommend that they undertake revisions of the text to clarify this as the purpose and result of this portion of the study. These revisions should affect the abstract, introduction, results, and conclusions. Specifically, the authors should make explicit in the text the fact that the MODIS GPP data set is modeled and includes meteorological inputs, making correlation likely between climate variables and GPP-based phenology. They should also clarify what their identification of linear relationships between GPP-based phenology and climate adds to previous understandings of these relationships.

 

In further reviewing the draft, I also noted some improvements needed in the Materials and Methods section. The Materials and Methods section contains subsection headings for several steps of the phenology identification process. However, it contains no subsection headings for the following elements of the methods, such as trend analysis. This should be added. Moreover, the methods section contains no description of how the correlation was computed among the phenology metrics and the annual GPP (results present in section 3.4). This text should be added to the Methods.

 

I look forward to seeing the evolution of this work as it represents a large and important undertaking toward better understanding our changing planet and its climate.

 

References:

 

Chen, Xiaona, Yaping Yang, and Jia Du. "Distribution and Attribution of Earlier Start of the Growing Season over the Northern Hemisphere from 2001–2018." Remote Sensing 14, no. 13 (2022): 2964.

 

Gao, Xuan, and Dongsheng Zhao. "Impacts of climate change on vegetation phenology over the Great Lakes Region of Central Asia from 1982 to 2014." Science of The Total Environment (2022): 157227.

 

Li, Linze, Xuecao Li, Ghassem Asrar, Yuyu Zhou, Min Chen, Yelu Zeng, Xiaojun Li et al. "Detection and attribution of long-term and fine-scale changes in spring phenology over urban areas: A case study in New York State." International Journal of Applied Earth Observation and Geoinformation 110 (2022): 102815.

 

Liu, Qiang, Yongshuo H. Fu, Zaichun Zhu, Yongwen Liu, Zhuo Liu, Mengtian Huang, Ivan A. Janssens, and Shilong Piao. "Delayed autumn phenology in the Northern Hemisphere is related to change in both climate and spring phenology." Global change biology 22, no. 11 (2016): 3702-3711.

 

Shen, Xiangjin, Binhui Liu, Mark Henderson, Lei Wang, Ming Jiang, and Xianguo Lu. "Vegetation greening, extended growing seasons, and temperature feedbacks in warming temperate grasslands of China." Journal of Climate (2022): 1-51.

 

Zhang, Rongrong, Junyu Qi, Song Leng, and Qianfeng Wang. "Long-Term Vegetation Phenology Changes and Responses to Preseason Temperature and Precipitation in Northern China." Remote Sensing 14, no. 6 (2022): 1396.

 

Zimmer, Scott N., Matthew C. Reeves, Joseph R. St Peter, and Brice B. Hanberry. "Earlier green-up and senescence of temperate United States rangelands under future climate." Modeling Earth Systems and Environment (2022): 1-17.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report (New Reviewer)

The figures can be further normalized, such as axis name and so on.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

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

This paper focused on the analysis of phenology changes of global vegetation and climatic drivers underlying these changes using the remote-sensing data and partial correlations for the period 2001-2020. The authors found significant advancement of SOS, slightly delay of EOS, and thus the lengthening of LOS. The climate factors attributed to the shifts in vegetation phenology, but with high spatial and temporal difference. The whole paper was generally clearly written and the results were principally sound. However, necessary verification and explanations are still indispensable in this study to solve readers' doubt about the accuracy of phenological metrics, the substantial improvements are still needed.

 

Main comments:

1.     Are the extracted phenological data reliable in global regions? How to verify the validity of the phenological data? These issues need to be clarified to improve the credibility and scientificity of the study.

2.     What is the formula for the FOD? How to detect DOY of SOS and EOS by FOD? I will recommend the author should be made FOD method details clear. Because DOY of SOS and EOS are the most important of data in this manuscript.

3.     I’m recommended to write a unified Pearson correlation coefficient formula.

4.     Whether phenology responds to climate change on a monthly scale or a seasonal scale.

 

Reviewer 2 Report

 

The manuscript presents an interesting study of coupling satellite remote sensing imagery with FLUXNET observations to systematically map the shift of SOS, EOS, and LOS in global vegetated area, and explore their response to climate fluctuations and feedback on GPP. They suggested that SOS advanced significantly and EOS was slightly delayed, which resulted in prolonged LOS in 72.5% of the vegetated area. The seasonal temperature and precipitation are attributed to the shifts in vegetation phenology. However, I do have some concerns about the methods and results. Please see my detailed comments.

I have a few major comments:

1)      How does the phenology calculation method deal with the regions with bi-modal shape growing seasons?

2)      How to test the robustness of phenology calculations

3)      RMSE and ME did not show direct evidence that the remote sensing phenology is correlated to flux tower GPP phenology

4)      More discussions about SOS-Temperature and SOS-Precipitation, EOS-Temperature, and EOS-Precipitation. I have concerns about some of these results.

L27, “resist” is a little bit weird word here.

L60-73, I understand the methods are important but seem not related to your results and discussions. You can simplify this paragraph and then put it in your method.

L78-80, is a pretty long sentence, maybe simply this sentence.

L75-93, this paragraph includes a lot of key points: climate fluctuations on phenology, Vegetation phenology trend changes due to climate fluctuations, shifted vegetation phenology’s feedbacks, and the correlation between vegetation phenology and productivity. I suggest separating it into at least two paragraphs and then giving more detailed introductions. Then give the limitations of these studies, and further state the objectives of this paper.

L113-116, simply this sentence, and provide the quality controls for the datasets.

L118, The study period of FLUXNET sites? Do you use all 212 sites for this study?

L146-L148, the tower is measuring the carbon dioxide flux, and then model the amount of carbon dioxide absorbed (GPP).

L155-l158, please be clearer about the inflection point detection method for phenology extraction

L163, please state that the authors have removed these pixels using quality control data.

L164, please cite the SG smoothing method.

L225, what is the Maximum composite?

L231, I am not convinced by “the relationship between GPP and LOS reflects the interactions between vegetation phenology and climate changes”

L234, 3.3., could you show the phenology timing difference between remote sensing and flux tower data, RMSE and ME are important, but they are not giving clear information if phenology from remote sensing and flux tower data are comparable.

L254-258, I did not see many SOS and EOS from March to May and after September in the South in Figure 4, and authors might need to state more details in the method of how to calculate Phenology differently in North and South. For example, the SOS in Australia is mostly from DOY 50 to 200. Could also put the DOY in the text so that readers can easily follow these figures.

L271-275, should the tropical forest have the longest LOS?

L315, Put uncertainties in figure 5.

L293, L299, L3030, where these values (0.19 days/year and 0.06 days/year, 0.24 days/year) can readers get from Figure 5?

L351-355, The temperature, and precipitation are monthly or seasonal?

L383. Figure 8, the plots in the figure could be larger, currently, it is difficult to tell the color difference.

L416 put the variations in figure 10.

L459, LOS, and GPP are significantly positively correlated in ~20%, not 76% so I am not sure if we can state “ enhancing effect on vegetation productivity and thus increased carbon uptake from the shifted vegetation  phenology”

 

 

 

 

 

 

Reviewer 3 Report

Thanks for giving me a chance to be a referee to review this manuscript. I am also glad to learn more about the global changes in vegetation phenology and processes. Overall, this is a rigorously designed study with a significant synthesis. Overall, I am fond of the idea that authors employ to associate vegetation phenology shifts with climate change and feedback processes.  It is a highly efficient approach and meets the aimed scope of Remote Sensing. Overall, I evaluate this is a good written manuscript with rare places that deserve minor changes. I have some gross comments and detailed suggestions that need authors pay attention to and modify to improve the quality of the study. I ask authors to modify the English in the article because I find some expressions of sentences may be not so native for an international readership. Thereafter, I may accept the submission for publication.

Title: Good.

Abstract: Apart from the phenometrics, amplitude of the seasonal variation might play an important role in vegetation productivity as it the product of both the length of season and maximum productivity in the season.

Introduction

The introduction is very good; the authors demonstrate a thorough knowledge of the published literature and highlight the importance and background to carry out this investigation.

Materials and Methods

Methods are technically strong and well explained. I am just curious about applying the derivative approach to derive the phenometric for each year. Would not it be a better approach to apply the derivative approach to the mean climatology of the GPP and then the corresponding threshold values, which could be indicator of physical cover, be applied to all the years to derive the SOS and EOS, just a thought.

Figure 2 is good.

Results

RMSE of ~one month, seems too much?

I wonder how the authors have dealt with the double and triple season crops, generally, there are 1-3 in year, and the cropping cycle of the crops might be different from forest and/other natural land covers, how did authors manage this problem?

Discussion:

I am not sure about the LOS of the Crops, should be discussed. Generally, should it not be significantly different from the forests. It might not be a good idea to merge all classes of forests into one class.

I am not sure, if the DOY never spans of two-years, generally phenology in some part of the world might include Dec Jan in a same cycle.

Conclusion

No Comments.

Reviewer 4 Report

Fang et al. derived vegetation phenology from MODIS GPP and land cover data using the first-order derivatives of GPP. They then validated the dates against those derived from FLUXNET and calculated the trends and the relationship between phenology and temperature, precipitation, and GPP. I think characterizing vegetation phenology at a global scale using satellite imagery is a good attempt to understand large-scale phenology variability, but the adopted methods can introduce large biases into the results and there are critical discussions missing on the uncertainties related to both the remote sensing dataset and the climate dataset. Below I have some general comments and specific notes. 

 

General comments:

1. My major concern is uncertainty. The authors adopted an 8-day composite of estimated GPP from MODIS and smoothed and interpolated the data to daily resolution. They then acquired the maximum and minimum first-order derivatives as the start of growing season (SOS) and end of growing season (EOS) dates. Since the trend and standard deviation of SOS and EOS are relatively small (~0.19 day/decade and ~10 days in the mid-to-high latitude regions), each step could have introduced large uncertainties and noises into the results. Though the authors provided a brief discussion on the first-order derivative method they used, they did not address the other uncertainties or discuss the potential influences of their choices of methods on the results in the manuscript. In addition, interpolating precipitation from 10km to 500m using a bi-cubic interpolation could result in large biases too. Given the large uncertainties, I am not sure how reliable the results and conclusions are. 

2. Related to point 1, grassland can have multiple green-up and senescence in some regions (e.g., some grassland ecosystems in North America have two peaks or two growing seasons during one calendar year), which may have also influenced some of the results in this paper.  

3. The reasoning seems circular when the authors discussed the influences of the length of growing season (LOS) identified from GPP seasonal cycles on GPP.

 

Specific notes:

Line 28: what did you mean by “atmospheric carbon uptake”?

Lines 118-129: which Fluxnet GPP did you use (e.g., daytime vs. nighttime, CUT vs. VUT)?  Does it make a difference? Also, does the dominant land cover type in the MODIS grid always agree with the flux tower site in Fluxnet? 

Lines 137-143: downscaling a climate dataset from 10km to 500m using bi-cubic interpolation is probably not a good idea, especially for precipitation. Could you be more specific about how you did that?

Lines 148-151: I understand what you mean, but please check your grammar here. 

I have a few questions regarding the methods detailed in Section 2.2: 

(a) Is the 8-day product based on the same “best value” approach as in vegetation indices (i.e., the highest cloud-free greenness value)? If so, how would that influence your results?

(b) Since you were using maximum and minimum first-order derivatives to determine SOS and EOS, how would curve smoothing and interpolation influence the results? That is, as both steps would change the shape of the curve, how did you account for that uncertainty?

(c) Grasslands and savannas can have two or more growing seasons within one year, does the algorithm always pick up SOS and EOS from the same peak or would that change? If so, how did you account for the changes caused solely by that?

Figure 2: comparing b and d, maximum and minimum first-order derivatives in b seem to be different from when SOS and EOS are in d, how did you account for that uncertainty?

Lines 218-220: why did you separate the months like this? I imagine the season would differ at different latitudes. 

Lines 235-239: I don’t think an average RMSE of ~20 days means “generally in agreement”.

Lines 253-275 and Figure 4: I think you excluded evergreen forests in your methods. If you only excluded tropical evergreen forests, please specify that. 

Line 254: An SOS in May seems early for Northern forests. 

Line 257: Isn’t that expected because they are closer to the equator?

Lines 266-267: I am not sure I understand what you are trying to say. 

Lines 268-270: I don’t think that is a fair comparison since there are fewer vegetated regions in the Southern Hemisphere and the land cover can be different as well. 

Figures S1-S2: do you know why there are abrupt changes from positive to negative influences in East Asia? Again, I am concerned about the uncertainty associated with the 8-day temporal resolution and the smoothing methods. How do you know that this is a signal instead of noises induced by uncertainties in the MODIS data?

Figures 6 and 7: what is the y-axis? 

Also, if I understand correctly from the Methods, here you presented the highest absolute values of the correlations between seasonal temperature/precipitation and the dates. Could you also show which seasons these correlations are in?

Line 373: I think you meant “Great Plains” instead of “Midwest”. Also, that is why I asked about multiple growing seasons over grasslands. The signals may not have been picked up from the same growing season. 

Section 3.4 and Lines 392-394: I am not sure the logic is sound. You characterized vegetation phenology from GPP seasonal cycles and then concluded that phenology may influence productivity. It seems circular reasoning. 

Line 396: Reference 41-42 are both regional studies, and are you sure your results agree with Ref 6 (in the overlapping period)?

Lines 405-406: again this seems circular since LOS is derived from GPP time series.