Thank you for the insightful comments, which we hope will improve the manuscript. We are here responding to them one by one, presenting first the reviewer’s comment in italics, followed by our response.

1. The authors compare down-scaled SIF with FLUXCOM GPP. Downscaled SIF using from GOME-2 can include two sources of error: I) GOME-2 retrievals are known to be somewhat less accurate than say OCO-2 and TROPOMI and II) The downscaling itself might introduce errors. Given that we have more than 2 years of TROPOMI data, I don’t understand why a simple test of downscaled GPP with “original” TROPOMI SIF data can be performed. This would help evaluate the robustness of the product used.

Such a comparisons are already available in the Duveiller et al. paper (https://essd.copernicus.org/articles/12/1101/2020/), where the downscaled SIF dataset is independently validated with OCO-2 SIF observations and a further comparison of downscaled SIF against TROPOMI data is provided. The paper shows that there is a high spatio-temporal agreement with the first TROPOMI retrievals and justified the use of this global high-resolution SIF (with a long temporal archive) in the current analysis.

As we mention in the manuscript, due to GOME-2 sensor degradation we chose not to extend the paper’s analysis beyond 2014. As such, the time period considered does not overlap with OCO-2 or TROPOMI and therefore could result in data consistency issues.

         2.   Please always provide the reference wavelength for SIF (which is wavelength dependent) and clearly state whether it was length-of-day corrected or not. 

This is a good point, thank you for raising it. We will add to the manuscript the lines :

(L146) The two retrievals have a spectral wavelength around 740 nm, and differ in the retrieval method…

(L148) A correction factor to convert the instantaneous SIF to the daily average is applied to both datasets to ensure comparability with estimates at different acquisition times ((Frankenberg et al., 2011b https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2011GL048738 ; Köhler et al., 2018a [https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2018GL079031])

L270 In order to explore the potential capabilities of SIF as an early indicator of stress across different type of vegetation type, the response of downscaled SIF to anomalies in a number of meteorological variables is analysed.

->

The response of length-of-day corrected downscaled SIF to anomalies in a number of meteorological variables is analysed.

         3.   The dataset by Koehler et al wasn’t used but that decision is not well motivated (or described). What “bias” are the authors talking about? Statements like these really need to be rigorous, right now it is rather sloppy. 

We agree that the wording here is quite sloppy and can be mis-interpreted. For nearly all of the figures we have also run the SIFPK analysis. However the difference is marginal and we wanted to avoid duplication in the paper (which would double the number of figures).

The decision to use the SIFJJ dataset references the findings of Gregory Duveilller et al. (https://essd.copernicus.org/articles/12/1101/2020/). The paper compares each of the two SIF retrievals (at both coarse resolution and downscaled) with OCO-2 SIF (at 0.05deg), in terms of both the agreement, correlation and bias. Figures 1 & 3 show that there is 1) a slightly higher level of agreement between downscaled SIFJJ with OCO-2 compared to downscaled SIFPK with OCO-2, 2) similar levels of correlation between the downscaled SIF and OCO-2 for both retrievals, and 3) a lower level of bias between SIFJJ and OCO-2 than SIFPK and OCO-2. The (un-downscaled) SIFJJ tends to have more noise than SIFPK, however the downscaling process smooths over some of this noise (‘The JJ retrieval, which is known to be noisier and to have a smaller bias than PK, benefited particularly from the downscaling procedure, probably due to the embedded spatial smoothing step’). The decision is therefore made to use the SIFJJ dataset due to the fact it has slightly less bias and the downscaling process reduces the noise.

We propose to rephrase our motivation and make greater reference to the Duveiller paper, and we will replac L155-L159 in the next version with (changes underlined):

(L155). Duveiller et al. shows that the downscaled SIFJJ dataset is found to have a slightly higher level of agreement with the OCO-2 validation data than the downscaled SIFPK dataset and so is primarily used in this paper, and is henceforth referred to as ‘downscaled SIF’ (or SIFDS). The higher agreement likely results from the spatial smoothing step of the downscaling process that benefited the noisier SIFJJ more than the SIFPK.

         4.   To me, there is some circularity in the interpretations. Most importantly, the authors state that: “Proving this technique at a global scale provides evidence for the use of high-resolution SIF in monitoring the resilience of local ecosystems to environmental  fluctuations, an area of growing importance as extreme weather events become more frequent and more severe“. This statement is far reaching but it is actually based on just a comparison with FLUXCOM GPP, which implies that FLUXCOM GPP has the same potential (and could be provided in near real time as well). Thus, it is unclear what SIF could do that FLUXCOM (or other pure remote sensing products) can’t. The interesting cases would be those in which the products disagree but the author’s statement is based on the agreement in the IAV between the two.

We think the key point of this part of the analysis is that:

1. SIF is measured independently from the meteorological variables (it is remotely sensed and not modelled)
2. The FLUXCOM GPP is modelled using remotely sensed data including temperature and water (normalized difference water index) inputs.
3. The SIF response to meteo fluctuations is more or less the same as the FLUXCOM GPP response, as evidenced by the study.
4. Therefore, we have a remote sensing proxy for GPP, independent of meteorological variables, which we demonstrate is sensitive to the meteorological fluctuations in similarity to GPP.
5. Therefore, hopefully, this early study shows the potential for the use of SIF in measuring global plant growth in response to meteorological fluctuations.
   
   
   
   

The value of the analysis is the demonstration that that downscaled SIF follows the pattern observed in the FLUXCOM GPP, and as such provides a global  (mostly) independent RS near-real-time observation.

We will make the following change for clarifying purposes (changes underlined):

L280 ‘For comparative purposes, the FLUXCOM GPP is also included, however it should be noted that the product takes several remotely-sensed climatic variables as input.’

->

The FLUXCOM GPP is also included in the analysis, though, as noted, the FLUXCOM GPP product takes several remotely-sensed climatic variables as input and so is not independent of the meteorological drivers. The inclusion of the GPP product enables a comparison with the SIFDS, giving insight into whether the SIF behaves as may be expected of an independent proxy for GPP.

L580 Proving this technique at a global scale provides evidence for the use of high-resolution SIF in monitoring the resilience of local ecosystems to environmental fluctuations, an area of growing importance as extreme weather events become more frequent and more severe

->

Proving this technique at a global scale demonstrates that high-resolution SIF responds to meteorological fluctuations in a similar way to FLUXCOM GPP. As such it has potential as a near real-time indicator of vegetation status that, unlike FLUXCOM GPP, is independent of meteorological variables.

We also agree that some of the wording is overreaching in it’s conclusions from the analysis. In particular the use of the word ‘resilience’ and elsewhere ‘monitoring environmental stress’. We would like to draw your attention to a softening of the wording in the meteo-analysis. Some of this is described elsewhere in the responses to other reviewer comments, including the following changes:

L27 and demonstrates the utility of SIF as a measure of environmental stress.

->

and explores the similarity of the SIF and GPP responses to meteorological fluctuations.

 

L469     In this context the study suggests that it is possible to use high-resolution SIF as a near-real time measure of the resilience of ecosystems to climate fluctuations

->

In this context the study suggests that it may be possible to use high-resolution SIF as a near-real time measure of the response of vegetation productivity to climate fluctuations

   

L490     This suggests the possibility of using SIF in the near-real-time monitoring of vegetation stress in reaction to environmental conditions

->

This suggests the possibility of using SIF in the near-real-time monitoring of vegetation reaction to environmental conditions  

L580     Proving this technique at a global scale provides evidence for the use of high-resolution SIF in monitoring the resilience of local ecosystems to environmental fluctuations, an area of growing     importance as extreme weather events become more frequent and more severe

->

Proving this technique at a global scale     demonstrates that high-resolution SIF responds to meteorological     fluctuations in a similar way to FLUXCOM GPP. As such it has potential as a near real-time indicator of vegetation status that, unlike FLUXCOM GPP, is independent of meteorological variables.

         5.   Some (if not all?) of the variables analyzed (VPD, radiation) are also included as driver variables for FLUXCOM. It is thus unclear whether we are learning something new. The authors could do the same analysis as in Figure 10 but for FLUXCOM-GPP as well to evaluate whether the drivers (or limitations) between the datasets are identical or not. Only then would we learn something in my mind, right now a lot of the analysis is somewhat phenomenological.

I think this follows from the comment above and we apologise if this doesn’t come across clear enough in the text. Hopefully the changes mentioned in the previous comment make this clearer. These variables are drivers for FLUXCOM GPP however the object of interest in the paper is SIF, and so we are learning that SIF, an RS product measured independently of these variables, behaves in a similar way to one of the best global estimates of global GPP.

As for figure 10, it is possible to repeat it with FLUXCOM GPP (and indeed we have looked at this), but we agree, we would not be learning something new as it is simply a confirmation of (a circular argument: GPP is modelled with meteo variables, then we see how GPP responds to changes in those variables). We would like to stress that Figure 10 is not the main aim of the analysis, but rather Figure 9 is, as this shows that independent RS SIF responds to meteo-fluctuations in a way that we would perhaps expect if SIF serves as somewhat of a proxy for GPP, with the implication that this is useful in detecting future plant growth in response to short-term meteo variations. SIF offers value here as it is a near-instantaneously measured independent RS product (not modelled), and the analysis (demonstrated in fig. 9) confirms that its response to meteorological fluctuations follows what we expect from theory.

Figure 10 is simply an addition to show where SIF suggests a given meteorological variable dominates. We are happy to remove the figure if it is not considered useful to the analysis of course, and we agree it is not so surprising and doesn’t reveal a new insight about how vegetation growth responds to meteo-fluctuations. It is also possible to create the GPP-version or a SIF-GPP difference but I am not sure it is so useful to the paper, as fundamentally the paper is about downscaled SIF. Hopefully the suggested changes to the text in the previous comment make the reasoning clearer for not recreating the figure for GPP.

• Some minor   comments:
• Line 54: Please cite some of the original works on SIF and   GPP as well (e.g. Joiner et al and Frankenberg et al).

Thank you for pointing this out. We will include the following in the introduction:

- (introduction para3) First observations of global and seasonal terrestrial chlorophyll fluorescence from space, J. Joiner et al. https://bg.copernicus.org/articles/8/637/2011/ 2011

- (introduction para3) Disentangling chlorophyll fluorescence from atmospheric scattering effects in O2 A-band spectra of reflected sun-light C. Frankenberg et. al https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2010GL045896 2011

- Note, we have also added this reference for an alternative downscaling approach via a neural net ~ L85 https://bg.copernicus.org/articles/19/1777/2022/ 

• Line 58: Frankenberg and Berry don’t really talk about water availability. Maybe rather about a lower dynamic range in SIF yield vs GPP yield once stress kicks in.

This comment refers to line 56 : ‘[SIF] is considered to be the mechanism developed by plants to respond near-instantaneously to rapid perturbations in environmental conditions of light, temperature and water availability’

Yes, the reference we cite talks about reduced SIF yield when undergoing ‘drought stress’ (Fig. 2 in https://www.sciencedirect.com/science/article/pii/B9780124095489106323?via%3Dihub shows simulated responses based on a parameterization with leaves undergoing or recovering from drought stress). But we agree drought stress isn’t a ‘rapid perturbation’. 

We will change the line to ‘... to respond near-instantaneously to rapid perturbations in the environmental conditions of light and temperature, with the SIF yield also dependent on biophysical conditions such as the concentration of the CO_{2}-fixing enzyme Rubisco and drought stress’

•     Line 79: Please add citations for those data-products

Thank you, we will add at this position:

Retrieval and global assessment of terrestrial chlorophyll fluorescence from     GOSAT space measurements LuisGuanter     et. al.

Filling-in of near-infrared solar lines by terrestrial fluorescence and other     geophysical effects: simulations and space-based observations from     SCIAMACHY and GOSAT J. Joiner1

Global     monitoring of terrestrial chlorophyll fluorescence from     moderate-spectral-resolution near-infrared satellite measurements:     methodology, simulations, and application to GOME-2}, Joiner et al.

Overview of Solar-Induced chlorophyll Fluorescence (SIF) from the Orbiting     Carbon Observatory-2: Retrieval, cross-mission comparison, and     global monitoring for GPP

Global     Retrievals of Solar-Induced Chlorophyll Fluorescence With TROPOMI: First Results and Intersensor Comparison to OCO-2 Philipp Köhler     https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2018GL079031 

• Lines around 156: Lower bias, higher level of agreement: Please be more concrete, this could be anything. It is important to differentiate absolute biases (scaling factors), which are trivial from worse agreement as seasonality is not well captured. Also, this statement shows that there is considerable uncertainty in GOME-2 itself, thus it would be important to know whether the authors would draw different conclusions if they had chosen another data product.

Please see the response to comment 2. regarding what is meant by bias here. Hopefully this clarifies the issue.

• Line 184: I really don’t understand why the authors are working at 0.05 degrees rather than just aggregating everything to the native FLUXCOM resolution. Is there any good reason to introduce potential interpolation errors. My guess is the reason is convenience but     please prove me wrong.

The native resolution of the FLUXCOM RS +METEO run is 0.0833 dd, so bringing the 0.05 dd resolution of the downscaled SIF to 0.0833 would also lead to potential interpolation issues. To avoid interpolation issues, we would do better to aggregate all to a common resolution of 0.25dd, but since the entire point of the downscaling operation is to be closer to the actual elements on the landscape, aggregating and smoothing the fine granularity that has been obtained by downscaling seems to be a missed opportunity. Furthermore, by construction FLUXCOM is bound to be already spatially smoother than SIF. We have thus decided to bring everything to 0.05 and tolerate the potential small interpolation effects that might occur in FLUXCOM.

• Line 237: What is true though is that if SIF is zero, there certainly is no GPP (but not necessarily the other way around). Thus, there is a biophysical reasoning behind that assumption. Maybe the linearity assumption is the one that could be questioned?

We agree that it is the linearity assumption that should mostly be questioned here: we don’t believe there is any evidence that SIF and GPP do scale linearly all the way to zero, and indeed we don’t think there is any reason to believe it should scale like this (especially when we are considering different scales e.g. leaf level effect, or reducing further to quantum level effects). Forcing through zero a relationship that does not evolve linearly to zero will cause us to mis-estimate the parameters in the area where the relationship is linear, or at least more closely linear. Also, as stated by the reviewer, it may not necessarily be correct that when GPP is zero SIF must also be zero (in which case even at macro canopy scale, the intercept may not be zero).

Additionally, as a broader statistical comment, there are also a theoretical motivations for why one should never (more or less) force a linear regressions through zero, particularly on measured real world data.

• Figure 3: The IAV correlations are surprisingly good. It would be VERY interesting to compare the SIF-GPP slopes derived intra-annually from those inter-annually.

Thank you for the suggestion. We will produce them and analyse them, and then incorporate the results in the text. 

• Figure 6: Please use higher resolution for the final version (or vector graphics)

Thank you, we will implement this in the update.

• Line 454: “high VPD correlates with high cloud cover”. I must be reading this wrong, it doesn’t make sense and the causality of the sentences here is somewhat strange. Large scale atmospheric dynamics drive cloud cover and humidity, hence also VPD, temperature and solar radiation. There are feedbacks but it reads as if VPD is in the driver's seat here, which it isn’t

This is relation to the statement here:

‘Strong correlation is noticeable in the SIFDS response to meteorological fluctuations, as can be seen clearly in equatorial rainforests in figure 9: high VPD correlates with high cloud cover, limiting the solar radiation arriving at the leaves and, naturally, reducing temperatures and affecting rainfall (and therefore soil moisture)’

What we are trying to say here is that - looking at Figure 9 - the SIF responds in a similar, consistent way to many of the variables (e.g. in arid croplands, there are similar magnitude shifts in T2M, SSR and VPD, - and a similar reverse magnitude shift in SM - or in equatorial rainforests VPD SSR and T2M respond with similar magnitude). This is likely due to the effect of the meteo variables being co-dependent and can’t be treated in isolation. Therefore we can’t just say in isolation ‘as one variable - e.g. VPD - increases, SIF/GPP increases’ because each variable is related to the others. For example an decrease in VPD might generally occur (through some dynamic atmospheric effect, the details of which are unimportant) concurrently with a decrease in temperature and so it is not possible to say which meteorological is really causing the change in SIF (is it the decrease in temperature, or the decrease in VPD driving the SIF)? In reality the variables are all inter-related.

Re-reading the statement now, we agree that it is quite messy and overreaching the original point (which is already captured by preceding sentences) and implying a broader statement on the nature of atmospheric dynamics in rainforests, which we didn’t intend.

We will substitute the line with:

‘Codependence between the atmospheric variables means that it is difficult to directly explain fluctuations in SIF via individual meteorological variables in isolation, for example, the correlation between warmer temperatures and high VPD, results in a similar SIFDS response…’

• Line     469: Again, this statement requires caveats.

Please see the response to comment #4 and see the response to Reviewer 2’s comment regarding the fact that we are perhaps over-stating the  results with the use of the words ‘measure of environmental stress’ and ‘measure of resilience’.

In this particular case, we will make the change:

L469 ‘In this context the study suggests that it is possible to use high-resolution SIF as a near-real time measure of the resilience of ecosystems to climate fluctuations’

->

‘In this context the study suggests that it may be possible to use high-resolution SIF as a near-real time measure of the response of vegetation productivity to climate fluctuations’

• Line 481: “Purity” maybe state “quality”?

We agree and will change it in the revised version.

• Line 503: Given the low dynamic range of tropical GPP, this is not surprising. So the question is whether the lower correlation is just due to the lower dynamic range in the presence of noise or something else?

Yes, we agree. And the fact that the correlation is suppressed both spatially and temporally suggests that it is an issue of noise, or, similarly, high uncertainty in the FLUXCOM model in the tropics. We make suggestions as to the reasons behind this lack of range (e.g. saturation of fapar or constraints in the FLUXCOM model). As we mention in the discussion, the greater dynamic range in SIF opens the possibility of it’s inclusion as a variable in the FLUXCOM machine learning model.

• Line     581: See above, these statements can’t be made without explicitly re-stating the assumptions or caveats.

Please see similar responses in comment #4 and on the theme of ‘environmental stress’ . For this line in particular, we make the previously noted change:

L580 ‘Proving this technique at a global scale provides evidence for the use of high-resolution SIF in monitoring the resilience of local ecosystems to environmental fluctuations, an area of growing importance as extreme weather events become more frequent and more severe’

->

‘Proving this technique at a global scale demonstrates that high-resolution SIF responds to meteorological fluctuations in a similar way to FLUXCOM GPP. As such it has potential as a near real-time indicator of vegetation status that, unlike FLUXCOM GPP, is independent of meteorological variables.’

Thank you for the important comments,and suggestions for improving the manuscript. Here we present the reviewer’s comment in italics, and our response:

1. Thank you for the important comments,and suggestions for improving the manuscript. Here we present the reviewer’s comment in italics, and our response:
   
   Main     comment – choice of the reference GPP dataset
   
   The authors have selected FLUXCOM GPP data (8-day & 0.0833º “Remote     Sensing” runs) as a benchmark for the evaluation of SIF ability to     indicate GPP. I can understand this choice, as the FLUXCOM dataset is well established in the community and has been tested in several projects over the last years.
   
   However, I also have strong concerns about whether the conclusions of the study would hold if a different remote sensing-based global dataset, or tower-based GPP data, were taken as a reference. For example, the values and variability of the SIF:GPP slopes discussed in Section 3 would surely be different if global GPP estimates from e.g. the FLUXSAT https://daac.ornl.gov/VEGETATION/guides/FluxSat_GPP_FPAR.html or the     https://data.nal.usda.gov/dataset/global-moderate-resolution-dataset-gross-primary-production-vegetation-2000%E2%80%932016 VPM products had been used as a reference, even if those two GPP products are also based on remote sensing data. Also, recent papers comparing TROPOMI SIF retrievals with tower-based GPP conclude that most of the vegetation types in North America can be grouped in 2-3 statistically-independent SIF:GPP linear models (see Li & Xiao     https://doi.org/10.1016/j.rse.2021.112748 and Turner et al.     https://doi.org/10.5194/bg-18-6579-2021), as opposed to the 12 independent groups proposed in this study.
   
   It would be great if the authors could provide further evidence of the robustness of their findings by comparing to additional GPP data sets, these being global remote sensing-based, tower-based, or both. For example, the global FLUXSAT GPP data set is provided at 0.05º and a daily time set, so it should not be too difficult to include it in this analysis. Showing that e.g. Table 1 roughly holds for other reference GPP data set are used would be an important proof of consistency for the study.

Whilst there are many different global GPP products (in addition to many different SIF products), we decided that doing many comparisons of these different datasets was beyond the scope of the paper. The main focus of the paper is to analyse downscaled SIF and thus FLUXCOM is chosen as the comparison GPP dataset as it represents the current state of the art in global GPP: it is a well-known and established reference within the biogeoscience community. Just to note, we will add the reference to the FLUXSAT GPP paper: https://doi.org/10.1016/j.agrformet.2020.108092 in the second paragraph of the introduction in the revised version.

Early on, we did have a look at including fluxnet towers in the analysis, and added the SIF-GPP distribution of fluxnet towers passing certain quality criteria (in order to allow for comparison with our KG-PFT breakdown) to figure 5. Unfortunately, there are simply not enough high quality and comparable globally distributed fluxnet towers over the different vegetation covers to make the analysis worthwhile, though it did serve as a sanity check for the data. We would be happy to provide this figure if it is considered interesting for the reviewers (only a few KG-PFT categories contain data points), but we do not believe it would be a valuable addition to the paper.

Regarding papers grouping the SIF-GPP relationships into a smaller number of categories, I would argue we actually demonstrate a similar thing. We start with 6 different vegetation covers (treating the 4 different climate zones separately) and we argue that these vegetation covers can be reduced to just 2-3 statistically independent linear models in each climate zone. Indeed, in the conclusions, we state that (with noted exceptions) the different species SIF-GPP scaling response (gradient) is similar, but with a distinction in the systematic potential (intercept) between woody/herbaceous:

L567 ‘For the most part, the gradient of the spatial SIFDS-GPPFX response is similar between differing vegetation types, with the exceptions of temperate deciduous broadleaf forests, continental needleleaf forests and, particularly, equatorial broadleaf forests. However, the GPPFX systematic potential for a given SIFDS observation displays more variation between species, with some divergence between woody and non-woody plants.’

Similarly in the abstract:

L15 ‘an analysis of covariance (ANCOVA) shows that the spatial response is similar between certain plant traits, with some distinction between herbaceous and woody vegetation, and notable exceptions, such as equatorial broadleaf forests, and continental needleleaf forests.’

See also the conclusions of the ANCOVA results section (the text of which is copied and edited later on in this comment).

The main consideration in our analysis is whether the SIF-GPP relationships hold for different vegetation types in similar conditions, and not whether the different climate groupings should also be reduced. Therefore we don’t technically propose 12 groups, we are simply comparing within climate groupings to reflect a study that controls better for climate variation, and propose 2-3 groups within each climate zone. Comparing in similar climate zones reduces the biases/variabiltiy introduced by differences in distribution and climatic conditions by comparing within similar zones. Therefore in reality we consider 6 vegetation covers and combine them into GRA/CRO and woody trees, but we identify in each climate a few exceptions  such as EBF (in equatorial climates), DBF (in temperate climates) and in continental climates the woody species behave quite differently to each other depending on what metric and cutoff we use (there is no statistically hard cutoff for these groupings unfortunately). We try to explore the behaviour of these exceptions further in the text. It is important to note that not all vegetation covers exist in sufficient numbers in each climate for a full analysis in each climate (and so combining climate would introduce a bias in our conclusions here). Our choice is trying to separate (as best we can) the effects of distribution and climate on the combination of vegetation covers, as opposed to a global comparison of vegetation cover (which would incorporate, say, tropical and temperate evergreen forests into one unit - which we see from their behaviour in our results would be quite unjustifiable). Section 5.3 contains more details about the nuances and the fact that it is difficult to statistically distinguish groupings - which we think the wider literature supports. 

It is also important to note that there is evidence that some vegetation covers can be divided further, for example differences in C3:C4 response https://doi.org/10.1029/2020GL087474 so distinguishing groupings might have a dependence on the scale of the analysis. For example, looking at a global scale there might be so much variation within a vegetation cover that it is difficult to distinguish their relationships, whilst at a scale that controls for this variation, e.g. canopy level, more patterns may emerge.

To make the similarities between vegetation covers clearer, and to emphasise that we are looking within climate zones only, we propose rewording the final paragraph of the ANCOVA section so that instead of listing the individual groupings, we just say herbaceous and woody in each climate zone, with the exception of Eq-EBF, Continental broadleaf, etc, for example rewording line 377+ (with underline denoting changes):

Overall, the ANCOVA analysis suggests that there is a large similarity in the scaling of the SIFDS-GPPFX response (i.e. the slope) between vegetation covers, with the major exceptions of temperate deciduous broadleaf forests, continental needleaf forests, and, particularly, tropical evergreen forests. In terms of the scaling of the SIFDS-GPPFX slope, these three vegetation covers may be treated as being reasonably distinct, with at least around 5% and up to 20% of the difference between slopes being attributable to the vegetation classification. Amongst the other species where the slope does not distinguish between veg etation covers so prominently (with generally less than 3% of the slope variation attributable to the vegetation categorisation), the intercept, and therefore the systemic difference between the linear relationships, loosely depends on whether the species is woody or herbaceous, with higher values for woody species. The difference in the SIFDS-GPPFX response between cropland and grassland is particularly minor. A caveat must be made that there are some exceptions to these generalisations, and there is no statistically concrete global distinction between groupings of vegetation covers.

These results offer loose quantitative support for the larger trends observed in figure 5, and demonstrate that whilst there are broad similarities in the SIFDS-GPPFX response between different vegetation types, there are still distinctions that can be made based on the background climate conditions. A loose, possible grouping of vegetation covers may be suggested within the climate zones, whereby equatorial regions feature: herbaceous (CRO+GRA), EBF and DBF groups; arid regions feature: herbaceous and woody (DBF+ENF) groups; temperate regions feature: herbaceous, woody (ENF+EBF) and DBF groups; and continental regions feature: herbaceous, DBF, ENF, DNF groups. This reduces the climate-vegetation categories for which we expect differing SIFDS-GPPFX responses from 18 groups to 12 overall. 

->

Overall, when analysing the the scaling of the SIFDS-GPPFX response (i.e. the slope) between vegetation covers within a climate zone, the ANCOVA analysis suggests that there are large similarities, with three major exceptions of temperate deciduous broadleaf forests, continental needleaf forests, and, particularly, tropical evergreen forests. In terms of the scaling of the SIFDS-GPPFX slope, these three vegetation covers may be treated as being reasonably distinct from others within that climate zone, with at least around 5% and up to 20% of the difference between the slopes being attributable to the vegetation classification. Amongst the other species where the slope does not distinguish between vegetation covers so prominently (with generally less than 3% of the slope variation attributable to the vegetation categorisation), the intercept, and therefore the systemic difference between the linear relationships, loosely depends on whether the species is woody or herbaceous, with higher values for woody species. The difference in the SIFDS-GPPFX response between cropland and grassland is particularly minor. A caveat must be made that there are some exceptions to these generalisations, and there is no statistically concrete global distinction between groupings of vegetation covers across all climate zones.

The results demonstrate that within a climate grouping there are broad similarities in the SIF-GPP response of the considered vegetation classifications, excluding three key exceptions. When accounting for differences in the intercept, a loose possible grouping may be suggested of herbaceous and woody vegetation within each climate zones, with the exceptions of equatorial-EBF, temperate DBF, and continental forests (which can be fully distinguished when the difference in the intercept is considered, or split between broadleaf and needleleaf if considering only the scaling). This reduces the climate-vegetation categories for which we expect differing SIFDS-GPPFX responses from 18 groups to 12 overall, with around three distinct groups in each climate zone, depending on the aggressiveness of the grouping.

We will additionally add the following lines and references in the 5.3 discussion

L543: ‘The universality of the SIF-GPP relationship with respect to vegetation groupings is in area of active debate’.

(Adding the references: https://www.sciencedirect.com/science/article/pii/S0034425721004685?via%3Dih as an extra reference about the difference between C3/C4 crops SIF-GPP relationship - and showing that there may be cases with more difference within PFTs than between them. And  https://bg.copernicus.org/articles/18/6579/2021/bg-18-6579-2021.html that others find two classes of vegetation)

L546 Indeed it may be the case that there are more differences within certain vegetation covers, than between vegetation covers, and this effect may depend on the scale of the analysis.’

L547 It is important to note however, that vegetation cover in the analysis may partially be a proxy for other factors or regional variables, such as background climate conditions and soil properties \citep{https://www.pnas.org/doi/full/10.1073/pnas.1216065111 a reference to support that some of the variation that cannot be explained by weather or plant function might be related to ecosystem dynamics (this is part of the reason we compare across climate zones)} 

Additionally, In answer to a separate question, we reproduced figure 6 to show the difference between FLUXCOM GPP and GPP estimated from the SIF scaling based on vegetation covers, without including climate zones. Whilst we don’t go too deep into the specifics (for example re-running the ANCOVA in the absence of climate zones), the fit is noticeably poorer, which suggests that there is some value in distinguishing between different climates. Please see the discussion in response to this request

• Other comments    
  
  Abstract: I would shorten the description of the implemented methodology and would add 2-3 lines summarizing the main results

We agree that it could be made more concise and will rewrite the text for the next version (including removing reference to ‘environmental stress’ here, as discussed in other comments).

• L64, 2nd equation (I miss equation numbers): please, state that this equation applies to instantaneous GPP and SIF, but a temporal sampling factor should be applied to account for the different temporal sampling in GPP and SIF (daily & all-sky for GPP, instantaneous &     clear-sky for SIF). Related to this, one could wonder to what extent     some of the features under analysis (SIF:GPP slopes, IAV, response to environmental factors) are not driven by this temporal sampling mismatch (see     https://www.sciencedirect.com/science/article/pii/S0168192321001222).

1) Thank you, I will add ‘instantaneous’ to line 64:

‘Rearranging the equations for instantaneous SIF and GPP fluxes:’,

2) It is possible to add a temporal sampling factor (e.g. e_{samp}) to the equation, but as the discussion is somewhat theoretical at this point, it is assumed the SIF and GPP are occurring simultaneously and therefore it is not required

3) Throughout the manuscript we are referring to daily averaged SIF and GPP values (see our response to Reveiwer 1’s  comments on this theme also, this will be made clearer in the revised text). Therefore, and feel free to correct, shouldn’t this principally result in an extra uncertainty in the measurement of the features, rather than being a driver of the features? We agree that there is going to be a significant clear sky bias however in these (and most) SIF studies however, and this may also explain some of the divergence seen in the tropics though. 

We agree that the recommended paper shows the value in correcting for all-sky conditions in the future studies and will discuss further in the next manuscript.

4) We will add equation numbers

• L76 “FLEX, scheduled for launch in 2023” – I think it will be at least 2025     https://www.esa.int/Applications/Observing_the_Earth/FutureEO/FLEX

We agree with this comment, and will update it in the revision.

• L258 “Downscaled FLUXCOM SIF” → “Downscaled GOME-2 SIF”

Thank you for spotting this typo

• Sec. 4.4 and Fig. 6: I think the “No climate zone” case (only vegetation types, without segmentation by climate zone) should be added, as this would represent the usual “PFT-based” scaling of other studies. It could include a test of how the SIF-based GPP     differs if only the two clusters proposed by Turner et al. are used (see major comment above).

Currently we provide the ‘no climate zone’ linear relationship parameters in figure 5 (these are the values on the left side - but these are for the vegetation types). However we can also apply these to Fig. 6 in order to add an ‘each vegetation category’. As you say, this would be useful for comparison with other papers, as well as comparing for comparing with our suggestion of 2-3 categories within a climate zone.

We add the figure to the next version as well as text on the additional panel and a short description . Note - we find that removing the climate zone grouping results in less agreement between the FLUXCOM GPP and the GPP estimated from the SIFDS-GPPFX vegetation relationships. The agreement is better than the case where only the climate zones are considered (and not the vegetation cover), suggesting that vegetation cover is more important than these broad koppen geiger climate groupings, but there is still a noticeable decline in agreement (compared to a full PFT-KG breakdown or the breakdown suggested by the ANCOVA analysis).

We propose adding the following line that discusses this to the text:

L409 When only the vegetation covers are considered, and no climate grouping is proposed, there is less difference between the estimated GPP and the FLUXCOM GPP than in the case of the climate groupings alone, suggesting that differences between vegetation covers are more important in determining the SIF-GPP relationship than the climate zone grouping. However there are still noticeable differences compared to the relationships that include a breakdown by climate grouping, as can be seen in the width of the inset histograms.

• Fig. 6, funny red line in the leftmost vertical label of the bottom left panel

Thank you for noticing this. We will adjust the future revision

• L453 (and elsewhere): “Strong correlation is noticeable in the SIFDS response to meteorological fluctuations, as can be seen clearly in equatorial rainforests”. - I think the rainforest case should be handled with caution, as a large fraction of the observed trends could just be due to signal issues and retrieval artifacts

This response also incorporates similar comments from Reviewer 1, and so it is worth seeing this comment also.

We agree that equatorial rainforests are a special case. The main aim of the comment was to give an example of where there is co-dependence in the meteorological variables, however it is a complex example, and the wording was confusing. We will instead substitute this line with:

‘Codependence between the atmospheric variables means that it is difficult to directly explain fluctuations in SIF via individual meteorological variables in isolation, for example, the correlation between warmer temperatures and high VPD, results in a similar SIFDS response…’

• Sec 4.5: the authors refer to this part of the analysis as an assessment of the response of SIF to “environmental stress”. However, I am unsure that the tiny signal of stress (subtle changes in LUE or photosynthetic pigments) can be captured by a downscale SIF product with a monthly sampling. Also, the acquisition time of the SIF data (morning for GOME-2, it would be midday for TROPOMI) will also play a role of their ability to indicate stress. I would recommend the authors to discuss these issues in the text.

We agree, and get the general impression though that we are perhaps over-stating the results with the use of the words ‘measure of environmental stress’ and ‘measure of resilience’ and should perhaps tone down the language a little. The aim is simply to convey that SIF may be useful as an indicator of plant growth in response to fluctuations in meteorological drivers and that lack of growth might indicate environmental stress. The results show that downscaled SIF responds to meteorological fluctuations in a similar manner to FLUXCOMGPP (which is dependent on the meteo-variables, whilst SIF is independently observed). Therefore we believe the results to be valuable.

Resilience is likely also the wrong word to use, and therefore we will remove reference to it and replace ‘environmental stress’ with less strong terms. I note the following changes for the updated version:

L20 With the demonstration of downscaled SIF as a proxy for GPP, the response of SIFDS to short-term fluctuations in several meteorological variables is analysed, and the utility of SIFDS as a measure of environmental stress explored.

->

With the demonstration of downscaled SIF as a proxy for GPP, the response of SIFDS to short-term fluctuations in several meteorological variables is analysed and the most significant short-term environmental driving and limiting meteorological variables determined.

L27 and demonstrates the utility of SIF as a measure of environmental stress.

->

and explores the similarity of the SIF and GPP responses to meteorological fluctuations.

L94 If downscaled sun-induced fluorescence is to be used as a proxy for ecosystem productivity or as a measure of environmental stress, it     is important to understand the spatial and temporal relationships

->

Removed ‘or as a measure of environmental stress,’

L106 the paper investigates the potential of downscaled SIF as a global     measure of environmental stress to fluctuations in several meteorological factors,

->

the paper investigates the response of downscaled SIF to fluctuations in several meteorological factors,

L270 In order to explore the potential capabilities of SIF as an early indicator of stress across different type of vegetation type, the response of downscaled SIF to anomalies in a number of meteorological variables is analysed.

->

The response of length-of-day corrected downscaled SIF to anomalies in a number of meteorological variables is analysed.

   

L469 In this context the study suggests that it is possible to use high-resolution SIF as a near-real time measure of the resilience of ecosystems to climate fluctuations

->

In this context the study suggests that it may be possible to use high-resolution SIF as a near-real time measure of the response of vegetation productivity to climate fluctuations   

L468 In this context the study suggests that it is possible to use high-resolution SIF as a near-real time measure of the resilience of ecosystems to climate fluctuations.

->

L468 In this context the study suggests that it is possible to use high-resolution SIF as a near-real time measure of vegetation change in response to climate fluctuations, as well as demonstrating where vegetation may be resistant to certain fluctuations.  

L490 This suggests the possibility of using SIF in the near-real-time monitoring of vegetation stress in reaction to environmental conditions

->

This suggests the possibility of using SIF in the near-real-time monitoring of vegetation reaction to     environmental conditions

L580 Proving this technique at a global scale provides evidence for the use of high-resolution SIF in monitoring the resilience of local ecosystems to environmental fluctuations, an area of growing importance as extreme weather events become more frequent and more severe

->

Proving this technique at a global scale demonstrates that high-resolution SIF responds to meteorological fluctuations in a similar way to FLUXCOM GPP. As such it has potential as a near real-time indicator of vegetation status that, unlike FLUXCOM GPP, is independent of meteorological variables.

Note, some comments also relate to Reveiwer 1’s comments on the wording overstating the results somewhat, so please see the responses to these comments. On the point about the acquisition time, noting other reviewer comments, we will explicitly and clearly state that the SIF is length-of-day corrected in the next version .

We thank the reviewer for the insightful comments, and welcome the tougher questions to both improve the manuscript and contribute to the general discussion within the SIF community. Please find below a response to all points raised one by one below.

•  Do comparisons among products tell us anything? 

  I am skeptical of analyses that compare products and interpret the results as containing empirical information or insights into their relationships. For instance, downscaled SIF is not SIF and FLUXCOM GPP is not GPP. To me it seems their relationships would be very sensitive to or determined by their respective errors. I do commend the authors for describing SIF as downscaled SIF and GPP as FLUXCOM GPP in the text and the figuresn, but does their relationship really tell us anything about SIF and GPP? Many other papers written with these products or similar products do not often make this distinction, and simply interpret downscaled SIF products as SIF and GPP products as GPP. So, thanks to the authors for being more diligent.
  
  On this note, I think it is a fair question to ask why a downscaled product was used - why not use the raw TROPOMI data. Or use them both and discuss how the results differ. If you repeated this analysis with gridded TROPOMI SIF data, would you get the same results? We have demonstrated ways to use ungridded (Doughty et al. 2019 PNAS) and gridded TROPOMI data (Doughty et al. 2021 JGR) for such analyses.

We agree that the downscaled SIF is not the same as the original SIF retrieval (nor the exact SIF emitted by the surface either), and that FLUXCOM GPP is but an estimation of GPP. Both have errors and shortcomings, as do all measurements and models in science. However, we still think that comparing them and examining their relationships can bring some insights about the underlying processes, or better, on where they are jointly well represented and where they don’t. We do not think these patterns would necessarily solely depend on the error structure. There is also quite some independence between both data streams, as SIF currently does not enter as a feature in FLUXCOM. Seeing where and when these match allows to identify zones of interests where more investigation is warranted. We will try to reinforce this message in the manuscript.

Regarding why we use a downscaled product instead of TROPOMI data, a first response is that since we wanted to investigate the interannual variability, using a downscaled product ensures more years are available. There is a bit of a historical reason for this too, as this work indeed started a couple of years ago when there were much fewer years of TROPOMI. Also, the idea was not necessarily to explore any SIF signal to GPP, but specifically the downscaled one. Including TROPOMI data would definitely be interesting. However, downscaled SIF is currently not available for the later years of TROPOMI because the input GOME2 retrievals are discontinued and they suffer from sensor degradation. While other datasets correcting for this might be available,this would still require a lot of cross-comparison and explanation. We fear all this would considerably lengthen an already long manuscript, and we thus consider it to be beyond the scope of the current study.

• Downscaled SIF products
  
  A couple of comments regarding the downscaled SIF products. This is certainly not intended to be a jab at the SIF-LUE product, but I think there are a couple of issues with most of these products that have not really been addressed yet.
  
  First is that have shown in my JGR 2021 paper that there is a very weak or often no correlation between VIs and SIF in the tropics, and I have found the same to be true when using TROPOMI SIF and TROPOMI surface reflectances. However, the downscaled products use VIs or surface reflectance, along with machine learning or environmental scalars such as we use in LUE models, to predict SIF.
  
  How sound is it to predict SIF with surface reflectance or VIs in the tropics when they lack a correlation? SIF is affected by physiological processes that do not affect leaf/canopy optical properties - so is it really safe to assume that we can use reflectances to predict SIF? This question is particularly important for the tropics since they are such a strong driver of annual and intra-annual GPP and XCO2.
  
  Second, do the downscaled products actually reproduce the SIF signal? The downscaled SIF products were produced before we had a sizable amount of TROPOMI data, but now we have four full years of TROPOMI data. Ideally, platforms with more coarse spatial and/or temporal resolutions (GOSAT, GOME-2, OCO2/3) would capture the seasonality of SIF in the tropics as observed by a near-daily observer like TROPOMI - but do we know that yet? And do their downscaled products reproduce the SIF signal, the VI signal, or something in between?

So here we might need to explain more clearly some particularities of the downscaled SIF approach we use. The downscaling is applied independently at every individual time step at which SIF is available using a relationship that is calibrated regionally. The downscaling is thus more akin to an unmixing process of the coarse signal, where the high spatial resolution explanatory data is trying to allocate where there is a higher likelihood to find higher values and where there might be lower values, and thereby try to best distribute these values around. However the actual signal that needs to be distributed remains that of the original SIF (gridded) observation. This is not the same thing as “predicting SIF with surface reflectance”, which is what other machine-learning downscaling methods try to do.  

Regarding the tropics, if we consider a homogenous land cover (e.g. tropical forests), and if we do not have any relationship between SIF and VIs over this area, this relationship will of course not help to disaggregate variation of SIF within that homogenous forest. Changes in LST or NDWI, which might have some relationship, could contribute partially despite a lack of relationship with the main VI (NIRv). But if no relationships are detected, no downscaling is done. The downscaled values retain the same value as the original SIF value at coarse spatial resolution. However, in reality, there will likely be some landscape fragmentation in which different land cover types can be detected at fine spatial resolution but not at coarse, and in these cases the downscaling (unmixing) approach should be able to disaggregate the coarse SIF signal based on differences detectable in NIRv (and LST and NDWI). This might not be sub-variations in SIF within the same landcover type, but it will be valuable to have a better relationships with different land cover types.

Finally, the downscaled (unmixed) product was benchmarked as suggested with respect to OCO2 observations in Duveiller et al 2020, and the agreement was good. Again, we do not pretend it “predicts SIF”, but that it tries to “disaggregates the GOME2 signal to a finer spatial resolution”.

• Analysis by land cover type
  
  Personally, I am not a fan of grouping land classes to investigate drivers of variables – in this case SIF and GPP. For instance, GPP in EBF in Africa or SE Asia can be driven by a different set of drivers than those in the Amazon. Even within the Amazon basin itself, there is a distinctive gradient in precipitation, temperature, VPD, etc. that is not static in space or time. Drivers of photosynthesis are determined locally by local environmental present and historical factors, disturbance history, species composition, human management, physiological processes, and many other local factors other than just land cover functional type.
  
  Thus, drivers should be investigated at the pixel level. Why not determine the drivers and their strengths and show it on a map? I am highly skeptical of any results that claim things like ‘GPP for this land cover type is driven by x’ or ‘SIF is driven by x for this vegetation type’.
  
  Also, the majority threshold used is somewhat subjective and arbitrary. Even at 75% majority land cover type, a sizable portion of the signal (GPP, SIF, or spectra) is driven by a land cover type other than the one you are interested in. Thus, there is an inevitable bias in the results that can’t be remedied. For instance, the seasonality in moist EBF of the Amazon is extremely subtle. Thus, even a small area of another vegetation cover type, such as crop or grassland, may dramatically alter the seasonality for a gridcell. Also, setting a 100% land cover threshold is unreasonable as one will end up with very few pixels for analysis, especially at 0.05-degree resolution.
  
  I have done these analyses myself while writing my 2021 paper published in JGR. I began the analysis by grouping by land cover type, but I obtained very different answers according to the majority % cover threshold that I used. I actually scrapped the entire paper and analysis in favor of showing the SIF-GPP and SIF-VI relationships at the gridcell level as maps, as it was not fair to extrapolate a relationship among all land cover classes globally as being characteristic for that land cover type when in reality the relationships and spatio-temporal relationships were much more complicated. And there are a lot of maps in that paper!
  
  

We agree, grouping by PFT has serious shortcoming regarding the simplification of the world and the reduction of local/regional particularities to an abstract grouping. In fact, this is why we are not only working at the PFT level but we also provide maps of continuous relationships between GPP and SIF (see figures 3, 6, 7). However, we also believe that grouping by PFT can still be useful, especially for some in the land surface modelling community who are still currently constrained by the PFT paradigm in order to calibrate their models. 

We will take the remark on board and make a more critical point of this in the revised manuscript and discuss more about the caveats of using PFTs.

• Would we expect the SIF-GPP relationship to be static?
  
  The SIF-GPP relationships shown in Figure 5 - wouldn’t we expect these relationships to vary over time and according to vegetation stress and other factors? Perhaps there is a seasonality to the relationships? What about a time series of their slopes, R2, or p values?

The relationships here are averaged over the full time period considered (2007-2014) and are growing season mean (i.e. more or less the annual mean). We take the mean GPP/SIF for each pixel over the time period. Then we do a pure spatial linear relationship. As such there is no evolution with time or time series. We do this to break down the SIF-GPP relationship into it’s spatial and temporal components separately. We also try to treat vegetation stress through the lens of fluctuations in meteorological variables in the paper. 

It would be an interesting extension to consider the evolution in time of the 8-day spatial SIF-GPP relationship, breaking down the seasonality, however we consider it beyond the scope of the paper. As discussed in a previous comment we will add maps of the slopes of the spatio-temporal relationships in the update, but we are also wary of overloading the, already large, manuscript with figures, for example by including R2 (in addition to the correlation) and p-values.

• Minor comments
  
  Land cover data – Is it really the case that the pixels selected had ‘no change’ over 2007 – 2014? Classification errors can cause estimated land cover fractions to change slightly from year to year. Also, land cover is changing. What was the threshold for no change? If a gridcell changed by 1%, from let’s say 95% to 96%, or vice versa, was it excluded? Or did you mean there was no change in the majority land cover classification?

Yes, this is a good point to raise. We are aware that land cover could change over the period and partly affect the results, but this would overall be considered negligible at the spatio-temporal aggregation scales we are considering. More specifically, what we meant is that there was no change to the majority land cover classification over the time period. We will clarify this in the text (bold):

L117: To ensure a high homogeneity in the selected data, the dominant vegetation type must cover at least 75% of a pixel and with no change in the majority land cover classification over the considered years, 2007-2014.

• Line ~40: There was a good LUE model review paper recently, see Yanyan Pei et al (2022).

Thank you, we will add this reference.

• Line 69-70 – This linearity can be said for spaceborne SIF sometimes, but certainly not tower or leaf-level SIF measurements. Youngryel Ryu showed that at the short term, SIF is more related to APAR, and the Marrs study and Helm study show that SIF and GPP can be decoupled at both short- and long-term scales and at both the leaf and canopy scale.

We agree and we will ensure to be clearer on stating that the linearity is only sometimes for and for spacebourne spacebourne measurements. We suggest the rephrasing (changes bold):

‘Indeed, there is a substantial body of evidence that shows that SIF, measured from space-based instruments, is positively correlated with leaf photochemistry, often exhibiting a generally linear relationship in both space and time, and across spatio-temporal scales

• Line 83 – The argument for not using TROPOMI is confusing. Apart from OCO2/3, it has the highest spatial resolution and certainly the highest resampling with near daily global coverage. GOME-2 is far inferior in these respects, so I recommend the authors better justify their use of GOME-2. There is certainly nothing wrong with the authors using their downscaled GOME-2 product, I am happy to see it.

TROPOMI is certainly the most interesting instrument, but it was lacking the temporal archive length to sample the multiannual variability, especially at the start of our study. The work will certainly be able to be continued later on with TROPOMI, but we deemed it was useful to show what could be done with a downscaled product of the sub-optimal GOME2 instrument. We will try to ensure the justification for not using TROPOMI here is clear.

• Downscaled product - Can you include your equations here? Also a brief description of your data sources for each variable would be helpful. From what I remember, the method follows  SIF = f(VI) * f(T) * f(W) using MCD43C4 (VIs), MYD11C2 (LST), MOD16A2 (ET).

We will add the details as requested. We had initially considered to remove them for brevity, but we now realize they help in the general understanding