Emerging negative impact of warming on summer carbon uptake in northern ecosystems

Most studies of the northern hemisphere carbon cycle based on atmospheric CO2 concentration have focused on spring and autumn, but the climate change impact on summer carbon cycle remains unclear. Here we used atmospheric CO2 record from Point Barrow (Alaska) to show that summer CO2 drawdown between July and August, a proxy of summer carbon uptake, is significantly negatively correlated with terrestrial temperature north of 50°N interannually during 1979–2012. However, a refined analysis at the decadal scale reveals strong differences between the earlier (1979–1995) and later (1996–2012) periods, with the significant negative correlation only in the later period. This emerging negative temperature response is due to the disappearance of the positive temperature response of summer vegetation activities that prevailed in the earlier period. Our finding, together with the reported weakening temperature control on spring carbon uptake, suggests a diminished positive effect of warming on high-latitude carbon uptake.

This may not affect their analysis, but I don't think that it can be neglected. I am not sure exactly how to consider this in their analysis, but perhaps looking at the difference between the zero-crossing and the minima is a better approach than looking at a fixed window over time. Of course this also presents a problem when looking at the fixed window of temperature data in your regression analysis. Further this analysis seems to assume that the temperature variance is stationary through time by looking at the 90 th percentile, but it is not shown that the variance is constant.
2.) It is interesting that there is such a renewed focus on the seasonal cycle originally identified by Keeling et al. in 1996 and yet nobody has attempted to solve the most intriguing puzzle identified by Keeling et al. that there was a strong positive relationship with temperature-but that seasonal amplitude lagged temperature by 1 to 2 years! Given the strong interannual variability in Fig. 1 of Wang et al. I think that if they did a lagged temperature analysis they would probably find a positive correlation with temperature in the preceeding years. While Keeling did not provide a conclusive explanation of this time lag it suggests complex feedback mechanisms with some memory effects. This previous relationships should at least be mentioned in the discussion.
3.) Barrow is located on the Arctic Ocean and is also influenced by potential changes in air-sea gas exchange. Figure S2 is misleading and suggests that this site is only influenced by land fluxes. While it may be safe to assume that air-sea gas exchange has very little impact on the seasonal cycle of atmospheric CO2 at other locations, this is probably not valid due to 'sea-ice phenology' that is changing over time.

Specific Comments:
L 38 This is true except for all the papers that have focused on changes in the amplitude of the seasonal cycle! L 60 this fixed window ignores phase changes in the timeseries (see Graven et al.) L 61 I don't know why it is becoming popular to summarize results in the introduction because I just read a summary of the results in the abstract. This section is redundant.
L 86 These changes in gamma results should be presented in Fig. 1 (or else in supplementary).
L 104 No arctic ocean in the footprint-impossible.
L 116 It sounds as though Wang et al. at least considered the phase shift in their analysis. However the caption to Figure S3 still implies a fixed window from DOY 178 to 235. What if you let this window vary based on the data. These results are possibly more important. Did they shift the temperature window accordingly in their regression analysis?
L 162 This assumes that temperature variance is not changing over time-not a valid assumption.
L 177 this is interesting because increased soil moisture could promote CH4 instead of CO2 losses L 183 I would restructure discussion and start with total ecosystem respiration (TER) and then move towards heterotrophic respiration (HR).
L 208 I suggest you go back and re-read the Keeling paper and discuss temperature sensitivity in a broader context Assume that air sea gas exchange plays little role in changing C uptake over time. (Fig S2) L93. What does "Main" summer footprint mean? 50% of the entire foot print? Be specific.
L203. Shifting to a significantly "negative state" is unclear. Rewrite the sentence. L206 Change to: to "increased" net carbon…" L210-213. Cite papers going back to IBP (1970s) and since including those in nature. L243. Performed "CO2 'measurement' analysis" is unclear. What is measurement analysis? Explain.
L306-309. What tower observations? How was this done? Where are the results? L354-355. What does this mean? How do you know this? How do you know that summer CO2 conc. Observed at BRW are mainly from terrestrial sources within the summer footprint area. Reference the SI and explain. How did you rule out ocean sources or sinks?
Reviewer #3 (Remarks to the Author): The manuscript demonstrates a significant change in the interannual relationship between summer land temperature and summer drawdown of atmospheric CO2 at high northern latitudes over recent decades. Careful comparison of the changing correlations between temperature and CO2 drawdown with proxy data for photosynthesis and ecosystem respiration indicate that the change in temperature response most likely results from heat and drought stress on plant productivity as a result of an increase in very warm weather. The data are from well-established measurement programs, the analysis is straightforward, the implications are important, and the paper is very well-written. I recommend publication with minor revisions as explained below.
The authors use a backward-in-time Lagrangian particle transport model (FLEXPART) to calculate upstream "concentration footprints" for the CO2 measurements at Point Barrow. The summertime footprint is used to define averaging areas for the various proxy data (NDVI, NPP, etc) that are used to investigate hypothetical mechanisms that might explain the change in temperature response of CO2 drawdown. This is a reasonable method and is based on well-established meteorological data and models, but it's surprising that the resulting footprint was localized to Alaska and eastern Siberia.
The authors cite studies by Gray et al (2014) and Zeng et al (2014) both of which attribute changes in high-latitude CO2 seasonality to increases in agricultural production in midlatitudes. A recent analysis by Barnes et al (2016) showed that CO2 seasonality at Point Barrow is strongly influenced by ecosystem processes in temperature latitudes, and that transport by baroclinic waves carries this seasonality into the high Arctic.
I suspect that the overall conclusions of the present manuscript do not depend sensitively on the footprint over which the proxy data are averaged. It seems likely that temperate ecosystems may also show changes in temperature response with recent warming. But the authors should address the issue of long-distance transport of seasonal CO2 variations from lower latitudes.
How sensitive are their conclusions to the particular footprint calculations they used? Perhaps 20 days of back-trajectory calculations are insufficient to track CO2 backward from Point Barrow to temperate croplands, pastures, and forests? At the very least the manuscript should cite recent work on long-distance meridional transport of CO2 seasonality and consider the implications for their results.

To Reviewer #1
[General Comment] In this analysis of "Emerging negative warming impact on summer carbon uptake in northern ecosystems" Wang et al. focus on summer carbon uptake at high latitudes and its changing relationship with temperature. Clearly the behavior of the global C cycle is changing over time and nowhere is this more evident than in the arctic where recent research has revisited changes in the seasonal cycle first identified by Keeling et al. 1996. Here Wang et al. conclude that summer C uptake is negatively correlated with summer average temperature and becoming increasingly more significant over time. While this analysis focusing on the peak of summer uptake is somewhat novel and potentially shows a decreased enhancement of photosynthesis due to temperature at high latitude, I have some concerns about assumptions made in analyzing this time series that is clearly changing through time.
[Response] Many thanks for your comments that help us to improve our MS. We carefully revised our manuscript following your comments and suggestions.
[Comment 1] By applying a fixed time window to their analysis the authors assume that the seasonal cycle has been stationary and that no phase shift has occurred in the time series. However, we know that the phase of the seasonal cycle at Barrow has shifted by 0.2 days/yr (Graven et al. 2013). Although this seems fairly trivial, this leads to a 15 day advance in the phase over a 30 year time series-thus in the Arctic June is the new July! Previous analyses have used the maxima and minima in the time series to estimate peak to trough amplitudes, instead of this fixed time window of summer carbon uptake. Therefore, it is conceivable that Wang et  This may not affect their analysis, but I don't think that it can be neglected. I am not sure exactly how to consider this in their analysis, but perhaps looking at the difference between the zero-crossing and the minima is a better approach than looking at a fixed window over time. Of course this also presents a problem when looking at the fixed window of temperature data in your regression analysis. Further this analysis seems to assume that the temperature variance is stationary through time by looking at the 90th percentile, but it is not shown that the variance is constant.
[Response] Following your suggestion, we performed the analysis using an inter-annually varying window approach, in which the summer CO 2 drawdown (SCD) is defined as the difference in CO 2 concentration between the climatological day of the year when the detrended CO 2 concentration crossed its long-term mean downwards (climatological spring zero-crossing date) and the climatological day of the year when detrended CO 2 concentration reached its annual minimum (climatological minimum date) ( Figure R1). Accordingly, the climate variables (including temperature, precipitation and cloudiness) for each year are averaged between inter-annual varying spring zero-crossing date and inter-annual varying minimum date across ecosystems north of 50°N.
As shown in Figure R1, both spring zero-crossing date and the minimum date advance at rates of 0.22 and 0.21 day yr -1 , respectively. The SCD increases at a rate of 0.065 ppm yr -1 and the corresponding temperature between zero-crossing date and minimum date increases at a rate of 0.04 °C yr -1 . By calculating the partial-correlation between SCD and temperature in the two periods, we again found that the correlation is not significant in the earlier 17 years (1979)(1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995) but is negatively significant in the recent 17 years (1996-2012) ( Figure R2). This result is in line with that based on the fixed-window approach, therefore suggesting that the varying phase of CO 2 seasonal cycle should not affect our main findings.
In order to test whether the temperature variance is stationary through time, we calculated the standard variation of daily temperature between July and August for each year. Our results indicated that there is no significant trend (P = 0.5) in the standard variation of summer temperature during the entire period  ( Figure R3), suggesting that the stationarity of temperature variance should not be regarded as an influencing factor in explaining the recent intensified negative temperature impact on summer carbon uptake. This result is also found if the standard deviation is calculated from summer temperature anomalies that are obtained by removing the low frequency signal from the daily temperature series base on the Empirical Mode Decomposition ( Figure R4).
In order to resolve the reviewer's concerns, we have changed "Second, to test whether this shift is due to the use of the month of August that includes a short time period during which CO 2 uptake declines ( Supplementary Fig. 1), we also computed SCD as the difference of CO 2 concentrations between the day of the year when detrended CO 2 crosses down its annual mean level and that when detrended CO 2 reaches its annual minimum. Again, we found a similar shift in R SCD-T from -0.28 (P = 0.32) during 1979-1995to -0.67 (P < 0.01) during 1996." into "We also explored the robustness of the result to the method used to calculate SCD. SCD was initially calculated as the difference in the CO 2 concentration between the first week of July and the last week of August in the detrended CO 2 record, but CO 2 uptake decreased briefly in August ( Supplementary Fig. 1), which could conceivably impact the results. Therefore, we calculated SCD using alternative methods. We first set SCD as the difference in CO 2 concentration between the climatological day of the year when the detrended CO 2 concentration crossed its long-term mean downwards (climatological spring zero-crossing date) and the climatological day of the year when detrended CO 2 concentration reached its annual minimum (climatological trough date). R SCD-T calculated from this different definition of SCD varied from -0.28 (P = 0.32) during 1979-1995to -0.67 (P < 0.01) during 1996 . 10 Fig. 11), so a fixed window for defining SCD using the climatological zero-crossing and trough dates may not be appropriate. We thus performed an additional analysis that allowed the spring zero-crossing and trough dates to vary interannually. We again found that R SCD-T shifted, from 0.42 (P = 0.12) during 1979-1995to -0.52 (P < 0.01) during 1996.
From this, one can conclude that the detected shift in R SCD-T is not an artifact of the method used to define SCD.". (Line 135-153 on page 7-8) In the Methods, we also changed "We computed summer CO 2 drawdown (SCD), which is adopted as an indicator of net summer CO 2 uptake, as the difference of CO 2 concentration between the first week of July and the last week of August. In addition, based on the detrended seasonal CO 2 curve, the mean estimate of minimum atmospheric CO 2 concentration and the day of year when CO 2 crosses down through its annual mean level during the period 1979-2012 at Barrow are around the day of the year (DOY) 178 and 235 ( Supplementary Fig. 1). Therefore, we also calculated SCD as the difference of CO 2 concentration between DOY 178 and DOY 235, which can effectively represent a period of strong CO 2 drawdown from the available CO 2 record." into "We calculated summer CO 2 drawdown (SCD), which was adopted as an indicator In addition, we have added the following text into Methods part.
"Before performing the correlation analyses, we calculated the standard deviation of daily temperature between July and August for each year and found no significant change during the entire period   (Supplementary Fig. 19), suggesting that the temperature variance is relative stable.". (Line 430-433 on page 22-23) Figure R1. The time series of spring zero-crossing date (a), the trough date (b), the SCD defined by inter-annually varying zero-crossing date and trough date (c) and the average temperature between inter-annually varying zero-crossing date and trough date (d). Figure R2. Same as Figure 1, but using SCD and corresponding climate calculated for the period between inter-annually varying date when CO 2 crosses down zero (inter-annually varying spring zero-crossing date) and inter-annually varying date when CO 2 reaches its annual minimum (inter-annually varying trough date).  analysis is adopted to decompose the daily temperature series to multiple frequencies.
We only contain the high frequency signals with the Hurst exponent lower than 0.5.
[Comment 2] It is interesting that there is such a renewed focus on the seasonal cycle originally identified by Keeling et al. in 1996 and yet nobody has attempted to solve the most intriguing puzzle identified by Keeling et al. that there was a strong positive relationship with temperature-but that seasonal amplitude lagged temperature by 1 to 2 years! Given the strong interannual variability in Fig. 1  [Response] Here we follow the reviewer's suggestion to conduct the lagged correlation analysis between SCD and temperature at the inter-annual timescale ( Figure R5). In order to well resolve the reviewer's concern, we have added the following discussion into the revised MS.
"The amplitude of the seasonal CO 2 concentration at Barrow was found to lag behind temperature by about two years 30 , and it was suggested that this lag was due to a lag in the response of net primary production to temperature. In contrast, SCD was not significantly correlated with summer temperature in the previous two years (or one year) for any of the study periods (1979-2012, or the two periods 1979-1995 and 1996-2012). This lack of lagged-correlation coincides with a non-significant lagged-response of summer productivity to temperature in the previous one or two years ( Supplementary Fig. 3). This result does not contradict the result from Keeling et al. (1996) 30 but it shows that if there is a lagged response of the peak-to-peak CO 2 amplitude to temperature, it is not due to a lag of summer CO 2 uptake.". (Line 94-103 on page 5) Figure R5. The inter-annual partial-correlation of summer CO 2 drawdown (SCD) (a), gross primary productivity (GPP) (b) and net primary productivity (NPP) (c) with summer temperature at various time lags (zero-year, previous one year and previous two years) for 1982-2011 (black), 1979-1995 (blue) and 1996-2012 (red) periods. °, * and ** indicates that partial-correlation coefficient is statistically significant at P < 0.1, P < 0.05 and P < 0.01, respectively.
[Comment 3] Barrow is located on the Arctic Ocean and is also influenced by potential changes in air-sea gas exchange. Figure S2 is misleading and suggests that this site is only influenced by land fluxes. While it may be safe to assume that air-sea gas exchange has very little impact on the seasonal cycle of atmospheric CO 2 at other locations, this is probably not valid due to 'sea-ice phenology' that is changing over time.
[Response] We agree with the reviewer that we need to consider the potential impact of changes in air-sea gas exchanges, since Point Barrow is located on the edge of the Arctic Ocean and its CO 2 variation could be influenced by ocean carbon fluxes according to our updated summer footprint area ( Figure R6).
As stated by the reviewer, changes in sea-ice phenology (melt onset and freeze-up) could induce variations in air-sea gas exchanges, potentially contributing to the variation of CO 2 concentration observed at Barrow station. In the context of global warming, the date of the onset of melting of Arctic sea ice advanced at rates around 2.8 day decade -1 during 1979-2004(Markus et al., 2009, which might stimulate summer air-sea CO 2 flux in the Arctic and potentially contribute to the variations of summer CO 2 concentration at Barrow. To test this hypothesis, we used sea surface temperature since which is closely correlated with percentage of open water in the Arctic during the summer (Galbraith and Larouche, 2011) and can then be regarded as an indicator for melt onset date. Our analysis shows that the correlation between SCD and Arctic SST north of 50°N are not significant in both periods ( Figure R7). A possible indirect effect of SSTs on the shift in R SCD-T was investigated using the partial-correlation between SCD and land temperature after controlling for the effects of cloudiness, precipitation and Arctic SSTs. Our results show that there is a similar shift in R SCD-T from 0.51 (P = 0.07) during 1979-1995 to -0.67 (P < 0.01) during 1996-2012 ( Figure   R8), tentatively suggesting the limited impact of changes in air-sea exchanges.
In order to address the reviewer's concern, we have updated Figure S4 to include potential emission sensitivity (PES) in ocean area. Furthermore, we have also added the following paragraph in the revised MS.  and SST (red) calculated as the average for July and August across ocean north of 50°N.
The inset illustrates the interannual simple correlation coefficient between SCD and Arctic SSTs. Figure R8. Same as Figure R7, but calculated partial-correlation between SCD and air temperature with controlling precipitation, cloudiness and Arctic SSTs. The lines are time series of anomaly of summer CO 2 drawdown (SCD, black) and temperature (red) calculated as the average for July and August across ocean north of 50°N. The inset illustrates the interannual partial correlation coefficient between SCD and Arctic SSTs.
* and ** indicates that partial correlation coefficient is statistically significant at P < 0.05 and P < 0.01, respectively.  Graven et al.) [Response] Following your constructive suggestion, we have added an additional analysis using an inter-annually varying window approach, in which the summer CO 2 drawdown (SCD) is defined as the difference of CO 2 concentration between the date when the detrended CO 2 concentration crossed its long-term mean downwards (spring zero-crossing date) and the climatological day of the year when detrended CO 2 concentration reached its annual minimum (trough date). Our analysis indicated that the results are in line with that based on the fixed-window approach, suggesting that the varying phase of CO 2 seasonal cycle should not affect our main findings. The details can be seen in the response of Comment 1 by Reviewer #1.
[Comment 6] L 61 I don't know why it is becoming popular to summarize results in the introduction because I just read a summary of the results in the abstract. This section is redundant.
[Response] We have changed "Here we analyzed the long-term record of atmospheric  19,20,21 , an ensemble of terrestrial carbon cycle models and simulations with an atmospheric transport model 22 ." into "The aim of this study was to understand the effect of temperature on summer CO 2 uptake in northern ecosystems and its decadal variation. We used the long-term record of atmospheric CO 2 concentrations from the Barrow atmospheric CO 2 monitoring station (71°N, 157°W, Alaska) 25 to calculate the summer CO 2 drawdown (SCD), which we used as an indicator of CO 2 uptake. SCD was calculated as the difference in the CO 2 concentration between the first week of July and the last week of August in the detrended CO 2 record (see Methods, Supplementary Fig. 1). Here, with simultaneous use of multiple satellite-derived products [26][27][28]  [Response] We have added the following figure to present gamma results in the supplementary. Figure R9. The inter-annual sensitivity of summer CO 2 drawdown (SCD) to summer temperature ( ) for 1979-1995 (blue) and 1996-2012 (red) periods. The summer temperature is calculated as the average for July and August across ecosystems north of 50°N (a), and the spatial average weighted by the potential emission sensitivities from FLEXPART over the vegetated land area within the multi-year mean summer footprint (b). * and ** indicate that partial correlation coefficient is statistically significant at P < 0.05 and P < 0.01, respectively.

[Comment 8] L 104 No arctic ocean in the footprint-impossible.
[Response] We have updated Figure S4 that now includes the ocean region. In addition, we have changed "Using the FLEXPART Lagrangian particle dispersion model 23 , we found that the summer footprint area of Barrow station is mainly restricted to regions of Siberia and Alaska ( Supplementary Fig. 2)." to "The area of the summer flux footprint of the Barrow station calculated using the FLEXPART Lagrangian [Response] In the prior MS, we calculated SCD as the difference of CO 2 concentration between the climatological trough date (day of year 235) and climatological spring zero-crossing date (day of year 178), which can effectively represent a period of strong CO 2 drawdown from the available CO 2 record. Since spring-zero crossing date and trough date could vary across years, we also calculate SCD using inter-annually varying spring zero-crossing date and trough date. In the partial-correlation analyses, the period used to calculate summer climatic variables (temperature, precipitation and radiation) is also inter-annually varying, and is consistent with that used to compute SCD. Our results are in line with that based on the fixed-window approach, suggesting that the varying phase of CO 2 seasonal cycle should not affect our main findings. The details can be seen in the response of Comment 1 by Reviewer #1.
[Comment 10] L 162 This assumes that temperature variance is not changing over time-not a valid assumption.
[Response] Before performing correlation analyses, we calculate the standard deviation of daily temperature between July and August for each year and find that there is no significant trend in the standard variation of summer temperature during the entire period (1979-2012) ( Figure R3 and R4), suggesting that the temperature variance is relatively stable through time. Please see response to Comment 1 for more details.
[ [Response] We strongly agree with the reviewer's concern on our English writing.
Thank you for your great patience in listing the problems that exist with the manuscript.
Following your suggestion, we have requested several fluent English speakers to play a more active role in the re-writing and editing of the whole manuscript. We hope that the revised manuscript could satisfactorily addresses all of your concerns.
To be more specific, we have changed "disturbances" into "fire disturbances".
[Comment 4] L33-34. This is debated. There are many reports of many northern areas switching from a C sink to an annual C source in recent decades. Review the literature staring in the 1990s (see e.g. Oechel et al. Nature 1993, 2000 to cite a few).
[Response] Following your suggestion, to be more accurate, we have changed "act as an important atmospheric CO 2 sink in the contemporary global carbon cycle" into "Arctic and boreal ecosystems play an important role in the global carbon cycle, and their carbon cycle responses to climate change become a major global concern 4,5 .".
(Line 38-39 on page 3) [Response] We thank the reviewer for bringing up this point. In this study, the summer footprint map, which is indicated by potential emission sensitivity from FLEXPART (or surface flux sensitivity from LMDZ), is used to weight climate variables (e.g. temperature). We then analyze the relationship between summer carbon uptake and the spatial average of weighted temperature. We did not set a cutoff value of sensitivities to select the summer footprint for the weighting, because a large fraction of the land surface had low sensitivity values. For example, values of potential emission sensitivity from FLEXPART above 0.1 s constituted 93.35% of the land-surface signal, and values above 1 s constituted 41.70% of the land-surface signal.
In order to well resolve the reviewer's concern, we have changed "a, b,  [Response] We have removed "profoundly" in the revised MS.
[Comment 13] L203. Shifting to a significantly "negative state" is unclear. Rewrite the sentence.
[Response] We have changed "Our findings provide evidence that the effect of temperature on summer CO 2 uptake in arctic and boreal ecosystems has been altered at the interannual timescale in the last three decades, shifting toward a significantly negative state, most likely due to the lowered effect of temperature on summer productivity (Fig. 3)." into "We demonstrated that the effect of temperature on summer CO 2 uptake in arctic and boreal ecosystems has been altered in the last three decades. A significant negative effect of temperature on summer CO 2 uptake has recently emerged, that is, warmer years coincide with less summer uptake, which is most likely due to the reduced effect of temperature on summer productivity (Fig. 3) [Response] Following your suggestion, we have included the early pioneer work (e.g. Oechel et al., 1995;Vourlitis and Oechel, 1999;Oechel et al., 2000;McFadden et al., 2003;Lund et al., 2010;Natali et al., 2010) in the revised manuscript. [Response] Following your suggestion, we have added "+" and "-" to indicate the positive and negative effect, respectively.
[Response] Sorry for our confusion expression. We have changed this sentence into: [Response] Sorry for this confusion. In this study, we did not directly use eddy-covariance measurements of CO 2 flux at the station (or tower) level, but used the global gross primary productivity (GPP) product (0.5° × 0. 5°, monthly, 1982-2011) that is generated by integrating a global network of eddy covariance sites, satellite remote sensing, and meteorological data in a machine learning algorithm. This global product has been widely used to understand the spatio-temporal dynamics of the global carbon fluxes, and to benchmark process-based land models.
To be clearer, we have changed "The second one is the upscaled gross primary productivity [Response] We agree with the reviewer's concern that Point Barrow is located on the edge of the Arctic Ocean and its CO 2 variation could be influenced by ocean carbon fluxes according to our updated summer footprint area ( Figure R6). To be more drawdown. This is a reasonable method and is based on well-established meteorological data and models, but it's surprising that the resulting footprint was localized to Alaska and eastern Siberia.
[Response] We have updated Figure S2 that now includes the ocean region. In addition, we have changed "Using the FLEXPART Lagrangian particle dispersion model 23 , we found that the summer footprint area of Barrow station is mainly restricted to regions of Siberia and Alaska (Supplementary Fig. 2)." to: Furthermore, the summer main footprint areas are further confirmed based on the adjoint code of the LMDZ atmospheric transport model ( Figure R10). According to our updated summer footprint area, Point Barrow is located on the edge of the Arctic Ocean and its CO 2 variation could be influenced by ocean carbon fluxes. We have also performed additional analyses to rule out the potential impact of changes in air-sea gas on our main finding (see details in Comment 3 by the Reviewer #1).
[ and 60 days back-trajectory calculations starting from the last week of August based on the adjoint code of the LMDZ atmospheric transport model. As shown in the Figure   R10, the main summer footprint area of Barrow station, based on the three different days of back-trajectory calculations, is restricted to regions of eastern Siberia, Alaska and its surrounding seas. We then calculated R SCD-T using climate variables averaged over the three different summer footprint areas, and found that R SCD-T shifted from a non-significant positive value to a significant negative value ( Figure R11), suggesting that the choice of particular number of days in back-trajectory calculation does not affect our main finding.
The analysis made by Barnes et al (2016) showed that CO 2 signals and seasonality can be transported from mid-latitude to high-latitude along surfaces of constant potential temperatures. They showed that the seasonal CO 2 amplitude at ~70ºN in most of the troposphere is more sensitive to seasonality of surface fluxes at ~30ºN than at ~60ºN.
While, a close examination of their Figure 11c indicated that the near-surface seasonal CO 2 amplitude (below 850hPa in their Figure 11c) is more sensitive to the seasonality of surface fluxes at 70ºN than from other latitudes (30ºN -60ºN). Their Figure 12 also showed that the seasonal amplitude of surface measurements along the isentrope of 265K is more sensitive to fluxes at 70ºN than at other latitudes. Since the Barrow   In order to resolve the reviewer's concern, we have changed "Using the FLEXPART Lagrangian particle dispersion model 23 , we found that the summer footprint area of Fig. 2). We therefore calculated R SCD-T using climate variables averaged over this footprint area (see Methods), and found that R SCD-T shifted from a non-significant positive value (R = 0.01, P = 0.97) to a significant negative value (R = -0.61, P < 0.05) (Fig. 1b) Fig. 5a-c). We also performed 40-and 60-day back-trajectory calculations using LMDZ and found that the footprint area for fluxes influencing the SCD did not change significantly compared to a 20-days influence ( Supplementary Fig. 5). R SCD-T between the two study periods still decreased when calculated using temperature spatially weighted with the footprint intensity. R SCD-T shifted from a non-significant positive value (R = 0.01, P = 0.97) to a significant negative value (R = -0.61, P < 0.05) in this test (Fig. 1b). R SCD-T also shifted when the footprint was derived from 40-and 60-day back-trajectory calculations rather than with a value of 20 days (Supplementary Fig. 6).". (Line 110-124 on page 6)

Barrow station is mainly restricted to regions of Siberia and Alaska (Supplementary
In addition, we have also added the following paragraph into the method section "Summer CO 2 uptake from the Barrow CO 2 data". "The second method used to define the summer footprint area was based on the adjoint code of the LMDZ atmospheric transport model 39  The authors have performed considerably more analyses to indicate that their results are robust. In particular, considering the effects of a temporally varying summer window of analysis and the effects of sea-ice. In considering these other factors the authors have concluded that their conclusions remain unaltered. However, in some instances it appears that using a temporally varying window lead to a change in the sign of correlation of temperature and SCD. In the instance of sea-ice it seems as though their correlations were slightly enhanced even though they looked at sea surface temperatures and not sea ice extent suggesting perhaps an interaction between sea surface and land surface temperatures. It should also be noted that the solubility of CO2 is decreased with increases in sea-surface temperature. Lastly, their analysis of diurnal temperature over time suggests that variance has not changed, but it was my understanding that minimum temperatures are increasing faster than maximum temperatures (see early work by Easterling) leading to a decrease in the seasonal range of temperatures. So the variance within any year should be going down at the global scale, but I am not sure if this is true in the Boreal/Arctic zone. Overall, this is a nice analysis.
Reviewer #2 (Remarks to the Author): I appreciate the work in revising this manuscript for Nature Communications. Many improvements have been made and are appreciated. The authors have done a better job citing previous work and being more careful in the phrasing and claims made.
The main point of this paper is stated by the title: "Emerging negative impact of warming on summer carbon uptake in northern ecosystems." However, the main data used in justifying the conclusions of the paper is atmospheric concentration data. The authors do not show the change, through time, in net summer uptake, and therefore cannot draw the conclusions that they put forward. For example, the authors conflate concentration and CO2 uptake: "summer CO2 uptake (the difference in CO2 concentration between the first week of July and the last week of August)" (lines 25-27). These are simply two different parameters and cannot be used interchangeably. Fluxes and concentrations can be affected by very different processes and source areas. Summer uptake, for the study region is never shown yet it is the basis of the paper. For this paper to move to be publishable, the authors need to calculate, validate, and present actual CO2 fluxes. To do that, they need to convincingly show that their calculation of net CO2 fluxes have a high degree of statistical confidence. They have not done that and this paper is therefore not appropriate for Nature Communications at this time. The authors attempt to justify the use of concentration and a surrogate for high latitude fluxes. However, if they want to make the claim that net summer uptake is becoming less sensitive to increasing summer temperatures, they need to work directly with fluxes, not concentrations.
One approach might be to refocus the paper on fluxes derived from satellite and model output. This would allow the authors to present and defend the calculated fluxes over the domain and discuss possible controls. The concentration data could be added for support or comparison, but not as key evidence in the paper.

Specific comments and observations:
L 31-34. Summer uptake is not shown. Net uptake is the result of ecosystem respiration and photosynthesis. Respiration is only known from model outputs and ecosystem carbon models for the Arctic are known to vary dramatically. Averaging a number of disparate model outputs does not necessarily give the correct resultant flux.
L 63-68. Summer CO2 drawdown maybe be an "indicator" of CO2 uptake. However, the authors have not made the definitive quantitative link between SCD and net CO2 uptake. This is a critical step missing from this analysis and presentation.
L 126-128. Randomly selecting 14 of 16 years of data seems pointless since almost all of the data is being used in each analysis. The full data sets should be used.
L 174. As stated in L 180, NDVI is only a proxy for productivity. They are not synonymous.
L 186-189. These are both indirect methods. The satellite derived NPP model has large uncertainties. The GPP data is poorly constrained by data.
L193-200. The number of hours when canopy photosynthesis is reduced by high canopy temperature has not been shown. Even on warm days, much of the day is below the temperature optimum for photosynthesis. Water stress might contribute to reduced photosynthesis. This is speculative as not data is shown.
L 209-212. Soil moisture and drought are two separate things. High temperature can cause drought even with high soil moisture. This is likely to occur, e.g., when root resistance is high due to low soil temperatures as occurs in Arctic soils.
L228-229. These models perform poorly against data. The mean or ensemble average of a number of disparate model outputs, one does not necessarily yield a true estimate of HR.
L231-234. If respiration continues to be depending on temperature, and if temperature continues to rise, it is unclear why this might not cause the purported decrease in summer NEE.
L236-238. Discussion. The statement that "We demonstrated that the effect of temperature on summer CO2 uptake in arctic and boreal ecosystems has been altered in the last three decades". There is no presentation of the actual rates of CO2 uptake over the last three decades. There is inference, but not verified data. Therefore, this statement cannot be made. Only statements about the change in rates of drawdown with summer temperature can be made. The paper overstates the conclusions possible.
L257. The statement that "High-latitude ecosystems have not shifted from a carbon sink to a carbon source" cannot be made base on this paper or on the one paper cited that uses inference. There is considerable literature and uncertaintly in this area, and it is insufficient to cite a single paper, especially one that is not definitive.
Reviewer #3 (Remarks to the Author): The revised manuscript addresses the concerns regarding transport from lower latitudes which I had with the original through some additional simulations using longer integrations of the Lagrangian parcel model. In addition, the authors have added some important context in the recent literature which were missing in the original manuscript.
In addition to my own concerns, the authors have done a nice job in addressing the concerns of the other reviewer, particularly with regard to aliasing of phase and amplitude effects in the timeseries.
I find that the revised manuscript is an important and timely contribution to the emerging understanding of carbon-climate feedbacks in high-latitude ecosystems, and recommend that it be published.

Reviewer #1
[ [Response] As stated by the reviewer, the sign of the correlation between summer CO 2 drawdown (SCD) and summer temperature during the earlier period becomes positive instead of staying negative if an interannually varying window was considered. But this positive value is not statistically significant at P < 0.05, and would then not overturn our main finding that the significant negative correlation between SCD and temperature emerged in the later period. To be more accurate, we have rephrased the sentence "We again found that R SCD-T shifted, from 0.42 (P = 0.12) during 1979SCD-T shifted, from 0.42 (P = 0.12) during -1995SCD-T shifted, from 0.42 (P = 0.12) during to -0.52 (P < 0.01) during 1996SCD-T shifted, from 0.42 (P = 0.12) during -2012 in that new test (Supplementary Fig. 12). From this, one can conclude that the detected shift in R SCD-T is not an artifact of the method used to define SCD. " as "In this additional analysis, we found that R SCD-T shifted from 0.42 (P = 0.12) during 1979-1995 to -0.52 (P < 0.01) during 1996-2012 ( Supplementary Fig. 12). During the earlier period, R SCD-T becomes positive instead of staying negative as it does when using the interannually varying window to define SCD. But the significant negative correlation between SCD and temperature still emerged in the later period leading to the main conclusion that this result is not affected by the method used to define SCD. " (Line 153-158 on page 8) In the second place, we strongly agree with the reviewer's point that Arctic sea surface temperature (SST) should not be equivalent to Arctic sea-ice extent (SIE), although which could be strongly correlated. We therefore performed partial correlation analysis between SCD and summer air temperature (T) (R SCD-T ) whilst controlling for the effects of summer sea-ice extent ( Figure R1). R SCD-T shifted from no-significant negative value (R = -0.16, P = 0.58) during the earlier period (1979)(1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995) to significantly negative one (R = -0.57, P < 0.05) during the latter period (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012). Our result emphasized that changes in summer Arctic sea-ice extent or Arctic sea-ice melt was not responsible for the shifted relationship between SCD and summer temperature. drawdown (SCD, black) and temperature (red) calculated as the average for July and August north of 50°N. The inset illustrates the interannual partial correlation coefficient between SCD and SIE. * and ** indicates that partial correlation coefficient is statistically significant at P < 0.05 and P < 0.01, respectively.
In order to resolve the reviewer's concern, we have then replaced "We would then expect a correlation between SCD and Arctic sea-surface temperature (SST), because SST is closely correlated with the percentage of open water in the Arctic during the summer 33 . We tested this hypothesis by calculating the correlation between SCD and summer (July-August) SSTs north of 50°N for the two study periods but found no significant correlation for either period ( Supplementary Fig. 13). A possible indirect effect of SSTs on the shift in R SCD-T was investigated using the partial-correlation between SCD and land temperature after controlling for the effects of cloudiness, precipitation and Arctic SSTs. R SCD-T shifted similarly to the original calculation ( Supplementary Fig. 14), suggesting that the impact of changes in air-sea exchanges on the SCD changes was limited." with "A possible effect of Arctic sea-ice melt on the shift in R SCD-T was then investigated using the partial-correlation between SCD and land temperature after controlling for the effects of cloudiness, precipitation and summer Arctic sea-ice extent (SIE). R SCD-T shifted similarly to the original calculation ( Supplementary   Fig. 13), suggesting that the impact of earlier Arctic sea-ice melt on the shift in R SCD-T was limited." (Line 164-169 on page 8-9) Lastly, to address your concern that asymmetric warming between minimum and maximum temperature could reduce the diurnal temperature range (DTR) and then influence the temperature variance, we calculated changes in DTR over the last 60 years for the arctic and boreal region (north of 50°N). The decreasing DTR mainly occurs in 1950s and 1970s. For the period 1979-2012, the summer DTR over the arctic and boreal region did not show any significant change. Therefore, the temperature variance is relatively stable in the boreal and arctic zone over the last thirty years.  One approach might be to refocus the paper on fluxes derived from satellite and model output.
This would allow the authors to present and defend the calculated fluxes over the domain and discuss possible controls. The concentration data could be added for support or comparison, but not as key evidence in the paper.
[Response] We agree with the reviewer that the use of actual net CO 2 fluxes in justifying our conclusion is straightforward. There are CO 2 flux products that might potentially be used in our analysis. Since large-scale direct observations of net ecosystem carbon exchanges are not available, the net CO 2 fluxes with spatially and temporally explicit maps are generally derived using the following three approaches: (1) the upscaling net CO 2 fluxes from a global network of eddy-covariance flux towers; (2) simulated CO 2 fluxes from ecosystem carbon cycle models; and (3) inverse estimates of net CO 2 fluxes inferred from atmospheric data using a global inversion of atmospheric transport.
First, the eddy covariance tower-based ecosystem carbon fluxes (Jung et al., 2009(Jung et al., , 2011 are upscaled from direct flux observations in machine learning algorithms, and have been used for ecosystem carbon cycle studies (Beer et al., 2010;Anav et al., 2013Anav et al., , 2015Bonan et al., 2011;Zscheischler et al., 2014). The data product generated spatio-temporal fields of carbon fluxes from the analysis of carbon fluxes and environmental drivers based on 224 flux tower sites (Jung et al., 2017). But the majority of eddy-covariance sites are located within the temperate climates. The sampling of boreal and arctic environments are sparse in the current eddy-covariance network (Williams et al., 2009). Moreover, in the generation of net ecosystem exchange (NEE), the impact of disturbances such as land-use change and fires are not considered (Jung et al., 2017). This estimate of NEE between the atmosphere and terrestrial ecosystems did not fully account for the carbon balance of ecosystems.
Second, ecosystem carbon cycle model is an alternative method to estimate large-scale net ecosystem carbon fluxes (Fisher et al., 2014a;Luo et al., 2015). Despite considerable progress over the past decade, ecosystem models still do not include or realistically simulate some processes such as permafrost carbon dynamics (Koven et al., 2011;McGuire et al., 2018), and fire disturbances (Bond-Lamberty et al., 2007;Beck and Goetz, 2011) over arctic and boreal ecosystems. Furthermore, there remains substantial uncertainties in simulating northern high latitude carbon cycle (Luo et al., 2012(Luo et al., , 2015Huntzinger et al., 2012;Lovenduski and Bonan, 2017). Large discrepancies exist not only in the direction, but also in the spatial distribution of net surface carbon fluxes (Xia et al., 2017;McGuire et al., 2012;Fisher et al., 2014b). For example, the inter-comparison of different models showed that the simulated arctic tundra CO 2 exchange for the period 1990 to 2006 ranges from 1 TgC yr -1 (almost carbon neutral) to a carbon sink of 255 TgC yr -1 (McGuire et al., 2012). We have also shown that most models of terrestrial ecosystems do not correctly reproduce the response of ecosystem productivity to temperature variation over the last three decades (see Supplementary Fig. 15).
Third, inversions of atmospheric CO 2 concentration provide a diver-down approach to estimate net surface CO 2 fluxes. The current relatively complete set of inverse model results, however, diverge on net ecosystem exchange (NEE) at high latitudes. McGuire et al. (2012) compared eight inverse models and found that inverse estimates of NEE ranged from a sink of 331 TgC yr -1 to a source of 173 TgC yr -1 for the period 1990 to 2006. Moreover, inverse estimates of net surface CO 2 fluxes, in particular the inter-annual variability, are sensitive to the composition of the CO 2 observing network used (Rödenbeck et al., 2003;Chevallier et al., 2012;Peylin et al., 2013). This would complicate our understanding of temporal changes in inter-annual temperature sensitivity of fluxes, based on inverse systems that generally used a continuously growing number of CO 2 observing sites throughout the whole time period (Peylin et al., 2013), or included dense CO 2 sampling from satellite observations during the post-CO 2 satellite period (Sellers et al., 2018).
The actual CO 2 fluxes derived from above-mentioned approaches could suffer from large uncertainties at high latitudes, and are therefore not well suited for our analysis. The long-term atmospheric CO 2 observation station has provided complementary monitoring of large-scale terrestrial carbon cycle from the atmosphere, since the seasonal rise and decline of atmospheric CO 2 concentration is due principally to the metabolic activity of terrestrial biosphere (Keeling et al., 1996;Randerson et al., 1999;Graven et al., 2013). A growing body of studies have used the current longest-running atmospheric CO 2 stations to understand temporal changes in regional-to-global carbon cycles and their associations with climate drivers (Wang et al., 2014;Graven et al., 2013;Forkel et al., 2016;Piao et al., 2017;Liu et al., 2018).
As the reviewer correctly pinpointed, concentrations and fluxes are two different parameters, since atmospheric CO 2 concentrations integrate net surface CO 2 fluxes from different regions through atmospheric transport. Although atmospheric CO 2 concentration is not a direct measure of terrestrial carbon fluxes, changes in CO 2 concentration at the Barrow station should be a good indicator for changes in terrestrial carbon fluxes over boreal and arctic regions. For example, previous studies identified that nearly all of the seasonality of CO 2 concentration at the Barrow site can be attributed to terrestrial carbon fluxes, with the largest contributions mainly from arctic and boreal regions (Graven et al., 2013). In addition, our transport simulations using the FLEXPART Lagrangian particle dispersion model also showed that the summer flux footprint of the Barrow station was mainly restricted to the regions of Siberia and Alaska (see Supplementary Fig. 4), implying that the CO 2 signal at Barrow station mainly reflects summer CO 2 fluxes from boreal and arctic regions.
Furthermore, we have performed an ensemble of transport simulations to demonstrate the linkage between change in summer CO 2 drawdown and that in summer terrestrial CO 2 fluxes.
We used historical net biome productivity (NBP) simulations during the period 1979-2012 from terrestrial ecosystem models, which participate in the historical climatic carbon-cycle model comparison project (Trendy) and the NBP dataset from MACC, to generate an ensemble of eight different NBP changes. We then use LMDZ atmospheric transport model to transform the simulated NEP into a point estimate of CO 2 concentration at Barrow station.
Our results showed that simulated changes in summer CO 2 drawdown (SCD) from the period 1996-2012 to 1979-1995 at Barrow station are strongly correlated with their corresponding summer NBP changes north of 50°N (R 2 = 0.65 , P < 0.01) ( Figure R3).
In order to respond to the reviewer's concern, we have added a section "The linkage between atmospheric CO 2 concentration and net surface CO 2 fluxes" into Methods of the revised MS.  Fig. 21). Although atmospheric CO 2 concentration was not a direct measure of terrestrial net CO 2 fluxes, change in SCD at the Barrow station is a good indicator for change in net surface CO 2 fluxes north of 50°N ." (Line 448-462 on page 22-

23)
We have also added the following sentences to stress the limitation of this study based on analysis of atmospheric CO 2 concentration at Barrow station. In addition, to be more accurate, we have changed the term "summer CO 2 uptake" into "summer CO 2 drawdown" in the revised MS. In the abstract, we have also changed "At first glance, summer CO 2 uptake (the difference in CO 2 concentration between the first week of  1996-2012 and 1979-1995. Note that the model LPJ was recognized as an outlier and is then not included in the correlation analysis.
[ [Response] As the reviewer correctly pinpointed, the use of the multimodel mean seems to provide a more robust estimate of heterotrophic respiration than any single model, but the accuracy of which might be questioned since no observation information is contained.
Moreover, these models do not include or realistically simulate some processes such as permafrost carbon dynamics (Koven et al., 2011;McGuire et al., 2018), and fire disturbances (Bond-Lamberty et al., 2007;Beck and Goetz, 2011) over arctic and boreal ecosystems. In addition, our results showed that most models do not correctly reproduce the response of ecosystem productivity to temperature variability over the last three decades (see Supplementary Fig. 15 in the revised MS). Because of a tight coupling between productivity and respiration, it is then questionable to use modeled respiration to quantify changes in the response of terrestrial respiration to temperature.
Here we therefore analyzed a global respiration dataset that integrates global soil respiration database (SRDB) (Bond-Lamberty and Thomson, 2010; version 3) in a climate-driven empirical model of soil respiration (Hashimoto et al., 2015;referenced as 'H2015' hereafter).
We applied the partial correlation analysis on the H2015 respiration dataset and found a significant positive correlation between H2015 respiration and summer temperature in both earlier (R HR-T = 0.85, P < 0.01) and later periods (R HR-T = 0.57, P < 0.05) ( Figure R4). This result suggests that change in the strength of the correlation would not be mainly responsible for the shift in R SCD-T . However, we should inform that this dataset still suffers from uncertainties (Bond-Lamberty and Thomson, 2010;Bond-Lamberty, 2018). Our understanding of the mechanism responsible for emerging negative temperature control on summer CO 2 uptake from the respiration perspective might be limited. Rigorous testing will require systematic and repeated measurements of respiration at larger spatial scales in the future studies. This limitation has been acknowledged in conclusion part of the revised MS (see also responses to Comment 1 by Reviewer #2).
Thanks to your suggestion, we mainly used SRDB dataset to explore change in the relationship between respiration and temperature in the revised MS. We removed the analyses related to respiration field based upon TRENDY models and MACC inversion system.  1979-1995and 0.72 ± 0.05 for 1996 . 18a) do not suggest a correlation becoming more positive in the later period. We also calculated heterotrophic respiration (HR) by subtracting satellite-derived NPP 27 from the MACC inversion net CO 2 flux and then calculated the correlations R HR-T . We obtained results similar to those for R RECO-T ( Supplementary Fig. 18b). We also used simulated HR indicates that partial correlation coefficient is statistically significant at P < 0.05 and P < 0.01, respectively.
However, the authors have not made the definitive quantitative link between SCD and net CO 2 uptake. This is a critical step missing from this analysis and presentation.
[Response] We strongly agree with the reviewer that summer CO 2 drawdown (SCD) should be used as an indicator of net CO 2 uptake. We have accordingly performed an ensemble of transport simulations to demonstrate a strong correlation between change in summer CO2 drawdown and that in summer terrestrial CO 2 fluxes north of 50°N (see responses to

Comment 1 by Reviewer #2).
We further performed LMDZ atmospheric transport model experiments to quantify the linkage between summer CO 2 drawdown (SCD) and net CO 2 uptake based on perturbations of surface fluxes. Specifically, we increased summer surface fluxes by a scaling value so that summer NBP north of 50°N increased by 1 PgC for each year of the period 1979-2012. For a given year, the scaling value for each pixel north of 50°N is calculated as the ratio of summer NBP to one unit of PgC.
To understand the uncertainty related to the choice of different NBP, we considered 7 modeled NBP from terrestrial ecosystem models participating in TRENDY project and the NBP dataset from MACC inversion system. Our results show that on average, one unit increase of PgC in terrestrial NBP north of 50°N can lead to an increase of 5.49 ± 2.70 ppm in SCD at the Barrow station ( Figure R5). We should inform that averaging an ensemble of modeled results does not necessarily give the correct sensitivity. The accurate estimate of this sensitivity relies on the accuracies of estimates of net surface fluxes and atmospheric transport model.
(Line 128-131 on page 7) [Comment 5] L 174. As stated in L 180, NDVI is only a proxy for productivity. They are not synonymous.
[Response] We have changed "so our first hypothesis is that summer plant productivity has become less positively responsive to temperature" into "so our first hypothesis is that summer vegetation activities has become less positively responsive to temperature". [Response] We agree with the reviewer that there are uncertainties in satellite-derived net primary productivity (NPP) and the upscaled gross primary productivity (GPP) data from a global network of eddy-covariance flux towers.
To further support our conclusion, we also used solar-induced chlorophyll fluorescence (SIF). SIF reflects an integrative photosynthetic signal from the molecular origin, and is increasingly used as a physiological-based proxy for GPP (Daumard et al., 2010;Frankenberg et al., 2011;Sun et al., 2017;Li et al., 2018). For example, Sun et al. (2017) used a high-resolution SIF dataset to show that there is almost a linear relationship between SIF and gross primary productivity at eddy-flux site in diverse biomes, further supporting the utility of SIF in revealing photosynthetic activities.
Here we used an extended satellite-derived SIF data during the period 2001-2012 (Zhang et al., 2018). This dataset is generated from a trained neural network that integrates surface reflectance from the MODerate-resolution Imaging Spectroradiometer (MODIS) and SIF from Orbiting Carbon Observatory-2 (OCO-2). The extended SIF dataset available for the all-sky condition has a spatial and temporal resolution of 0.05 degree and 4 days, respectively.  Fig. 14)." (Line 187-190 on page 9-10) We have added the following text related to SIF into "Methods" section.
"Solar-induced chlorophyll fluorescence (SIF) that reflects photosynthetic signals from the molecular origin is increasingly used as a physiological-based proxy for gross primary productivity (GPP) [53][54][55][56]  In addition, since satellite-based NPP could suffer from certain uncertainties, we have then  [Response] The actual high temperature stress that plants experience needs the definition of photosynthetic optimum temperature. However, most of our knowledge about the optimum temperature stems from site-level measurements (Niu et al., 2008;Way and Yamori, 2014).
We still lack essential knowledge about optimum temperatures at broad spatial scale.
Therefore, in this study, we defined extreme warm events that are potentially unfavorable for photosynthesis through counting the number of days or hours with daily or hourly mean temperature exceeding a certain threshold. We should inform that this approach is just used as an approximation to indicate the potential risk of high temperature stress on photosynthesis.
Following your suggestion, we have also calculated changes in number of extreme warm hours. For each day of the summer season (July and August), we extracted the maximum temperature and its two neighbors based on 3-hourly temperature data from ERA-Interim data (http://www.eu-watch.org/gfx_content/documents/README-WFDEI.pdf). These temperatures occurring throughout the entire study period 1979-2012 were sorted in ascending to determine the threshold value for the 90 th percentile. The number of hours with temperatures exceeding 90 th percentile were then referred to as extreme warm hours. The main conclusion is robust to the use of sub-daily temperature records in defining extreme warm temperatures. For example, the number of extreme warm hours increased during the later period ( Figure R7a). The patterns of the changes of R NDVI-T were roughly consistent with those of the number of extreme warm hours, particularly Alaska and eastern Siberia constituting the main footprint area of summer CO 2 changes at Barrow (Supplementary Fig.   4). NDVI in these areas had a significant negative partial correlation with the number of extreme warm hours when the data were statistically controlled for the effect of mean summer temperature ( Figure R7b).
In order to resolve the reviewer's concern, we have added the following sentence into the revised MS and Figure R7 into the Supplementary text.  root resistance is high due to low soil temperatures as occurs in Arctic soils.
[Response] As the reviewer stated, water stress could adversely affect rates of photosynthesis due to restricted CO 2 diffusion into the leaf resulted from stomatal closure, and inhibition of photosynthetic metabolism (Lawlor and Cornic, 2002;Flexas et al., 2006;Chaves et al., 2009).
Plant water stress is a combined function of soil water supply and atmospheric demand for water. Our analysis of soil moisture changes showed that summer soil moisture were slightly higher during the later than the earlier period (see Supplementary Fig. 18), suggesting that soil water supply is unlikely to cause water stress to plant growth.
Increased temperature could induce or exacerbate plant water stress through increasing atmospheric deficits to the degree which plants either lose water at a faster rate or close stomata. Here we also analyzed changes in the atmospheric vapour pressure deficit (VPD), which is indicative of atmospheric demand for water, between the earlier and later period. Our results showed that atmospheric demand for water generally increased over the main footprint area of summer CO 2 changes at Barrow ( Figure R8a). But NDVI in these areas had either a positive or non-significant negative partial correlation with VPD changes when the data were statistically controlled for the effect of mean summer temperature ( Figure R8b). It suggested that increased atmospheric demand for water over high-latitude ecosystems has a relatively low probability of imposing water stress constraints on plant growth. Therefore, changes in VPD might not account for the decrease in R NDVI-T .
As the reviewer mentioned, soil moisture and drought are two separate things. We have therefore changed "These values of soil moisture in summer were slightly higher during the later than the earlier period ( Supplementary Fig. 17), indicating no increase in summer drought. Thus, changes in soil moisture alone did not account for the decrease in R NDVI-T ." into "These values of soil moisture in summer were slightly higher during the later than the earlier period ( Supplementary Fig. 18), suggesting that changes in soil moisture alone did not account for the decrease in R NDVI-T .". (Line 216-219 on page 11) In addition, we also added the analysis of VPD change in the revised MS.
"In addition, increased temperature could induce plant water stress, through increasing atmospheric deficits, to such a degree that plants either lose water at a faster rate or close stomata. We also analyzed changes in the atmospheric vapour pressure deficit (VPD), which is indicative of atmospheric demand for water, between the earlier and later period. Our results showed that atmospheric demand for water generally increased over the main footprint area of summer CO 2 changes at Barrow (Supplementary Fig. 19a). But NDVI in these areas had either a positive or non-significant negative partial correlation with VPD changes when the data were statistically controlled for the effect of mean summer temperature ( Supplementary Fig. 19b). This result suggests that increased atmospheric demand for water has a relatively low probability of imposing water stress constraints on plant growth over high latitude ecosystems. Therefore, change in VPD should not be the main cause of the decrease in R NDVI-T ." (Line 219-230 on page 11) Figure R8. Spatial distribution of changes in vapor pressure deficit (VPD) during July and August (a) and the linkage between summer NDVI and VPD (b). The VPD changes are the difference between the 1996-2012 and 1982-1995 periods. The R NDVI-VPD is defined as the partial correlation coefficient of NDVI against VPD, whilst controlling for the effect of summer temperature. The gray dots indicate that the partial correlation coefficient is statistically significance (P < 0.05).
[Comment 9] L228-229. These models perform poorly against data. The mean or ensemble average of a number of disparate model outputs, one does not necessarily yield a true estimate of HR.
[Response] Thanks to your suggestion, we mainly used SRDB dataset to explore change in the relationship between respiration and temperature in the revised MS. We removed the analyses related to modeled respiration (see details in response to Comment 2 by Reviewer #2) [Comment 10] L231-234. If respiration continues to be depending on temperature, and if temperature continues to rise, it is unclear why this might not cause the purported decrease in summer NEE.
[Response] As the reviewer proposed, the finding that terrestrial respiration continues to be depending on temperature suggested an increase of temperature during the later period would continue to stimulate terrestrial respiration and thus decrease net carbon uptake. In this study, we are mainly concerned with change in inter-annual correlation between (indictor of) summer net carbon uptake and temperature in the last 3 decades, instead of change in (indictor of) summer carbon uptake itself. To explain the weakened temperature dependence of (indicator of) summer carbon uptake, we proposed and tested the following two hypotheses: a weakened temperature dependence of (indicator of) plant productivity and a loss of inter-annual correlation between terrestrial respiration and temperature.
Thanks to your suggestion, we have changed "We tested the hypothesis that the increased response of respiration to temperature was responsible for the change in R SCD-T ." into "We tested the hypothesis that the loss of temperature dependence of terrestrial respiration was responsible for the change in R SCD-T ." (Line 234-236 on page 12).
In addition, we have also changed "These analyses suggested that respiration continued to respond significantly positively to temperature in both study periods, but that the response was not stronger during the later period, implying that the change in R SCD-T was not due to changes in the response of terrestrial respiration to temperature." to "Our analysis suggested that respiration continued to respond significantly positively to temperature in both study periods, albeit that the response was not stronger during the later period, implying that the change in R SCD-T was not due to changes in the response of terrestrial respiration to temperature.". (Line 242-245 on page 12)