The study site (37∘35.75′ N, 100∘00.47′ E, 3571 m a.s.l.) is located in Gangcha County, Qinghai Province, in the north-eastern part of the QTP (Fig. 1a). The mean annual temperature and precipitation measured at the Gangcha National Weather Station were 0.1∘ and 389.4 mm between 1982 and 2011, respectively (Zhang et al., 2016). The site area has seasonal permafrost featuring frozen soils between January–March and November–December for a total of about 125–135 d (Zhang, 2014). The dominant species of the alpine swamp meadow ecosystem is Kobresia pygmaea, followed by Saussurea pulchra, Polygonum viviparum, and Potentilla saundersiana. The average plant height at the experimental site is 7.4 ± 1.5 cm, with a 97 ± 2 % coverage. Our study defined the growing season as the period between June and September. The early (June–July) and late (August–September) growing seasons, as they will be henceforth referred to, correspond with the early growing season and late growing seasons, respectively.

An eddy covariance (EC) system was installed at the study site (Fig. 1b) to measure the CO2 fluxes at a sampling frequency of 10 Hz from 2014 to 2018. Data for 2016 were missing due to equipment malfunction. The EC system included an open-path CO2/H2O infrared gas analyser, which quantified fluctuations in water vapour and CO2 fluxes. A 3-D sonic anemometer was installed at a 2.0 m height above ground to directly measure horizontal and vertical wind velocity components (u, v, and w). C flux data were recorded with a data logger (Campbell Scientific Inc.). An automated meteorological station was installed near the EC station to measure meteorological variables such as air temperature (Ta; ∘C), precipitation (P; mm), net radiation (Rn; W m-2), wind speed (WS; m s-1), wind direction (WD; ∘), relative humidity (RH; %), and vapour pressure deficit (VPD; hPa). The meteorological data were collected at 1 min intervals and subsequently resampled at 30 min time steps to keep pace with the EC data. More details on the in situ instrument specifications are summarized in Table 1.

Soil water content (SWC; %) was measured at depths of 10, 20, 40, 60, and 100 cm from the soil surface with EC-H2O sensors (Decagon Devices, USA) at a 10 min recording frequency. The precision of the EC-H2O sensors for soil moisture measurements was ± 0.03 m3 m-3. As the roots of Kobresia meadows are mainly distributed within the top 20 cm of soil, we focused only on the variation of SWC in the top 20 cm of the soil.

The half-hourly NEE data were calculated using the EddyPro software (version 5.2, LI-COR) from the 10 Hz raw data. During the calculation, three-dimensional rotation was used to correct the data by removing the effects of instrument tilt irregularity on airflow (Wilczak et al., 2001). Webb, Pearman, and Leuning (WPL) (Webb et al., 1980) correction was applied to calculate the averages of CO2 covariance, rectifying the air density variations induced by heat and water vapour. The half-hourly flux data were quality-checked based on several filtering algorithms, including (1) the rejection of outliers in sonic temperature, water vapour density, and CO2 density (Li et al., 2008; Liu et al., 2011); (2) the elimination of data 1 h before and after precipitation events (Li et al. 2008; Liu et al., 2011); (3) the removal of negative NEE during the non-growing season (from November to March) (Cao et al., 2017; Qi et al., 2021) attributed to the self-heating effect from EC instruments (Cao et al., 2017); and (4) the exclusion of measurements with weak turbulence conditions at night-time. The weak turbulence periods were identified by bootstrapping friction velocity (u∗) thresholds, as described by Papale et al. (2006). This approach effectively divided the data into 4-year and seven-temperature subsets with similar micro-meteorological conditions (except for u∗). The u∗ thresholds (5 %, 50 %, and 95 % of bootstrapping) were calculated specifically per year and temperature subset.

Based on those different subsets, we gap-filled and partitioned NEE (into GPP and Re) to spread the uncertainty variability that emerged from the different u∗ thresholds, similar to López-Blanco et al. (2017). All missing data were marked as -9999 (no data). Negative and positive NEE values represent the sink and source of C, respectively. Additionally, a standardized mechanism to fill NEE gaps is needed for adequate data processing (Moffat et al., 2007). Therefore, this study adopted the method described by López-Blanco et al. (2020) using the marginal distribution sampling (MDS) algorithm in the REddyProc gap-filling tool (Reichstein et al., 2016), which was readapted from Reichstein et al. (2005). Finally, NEE was separated into GPP and Re by applying the REddyProc partitioning algorithm (Reichstein et al., 2016) for further analyses. This partitioning method is based on the exponential regression of night-time respiration with temperature using the Lloyd–Taylor function (Lloyd and Taylor, 1994). Night-time periods were selected via current combined solar radiation and potential radiation thresholds based on the exact solar time, latitude, and longitude. REddyProc estimates the temperature sensitivity from a short-term period, and based on this short-term temperature sensitivity, it estimates the reference temperature in the continuous period of the entire dataset. These estimates were then used to calculate Re for daytime and night-time, while GPP was estimated based on the difference between NEE and Re.

In order to characterize how environmental conditions impact diurnal and seasonal variability of NEE, GPP and Re at this alpine swamp meadow, we used a novel method based on machine learning. Random forest (Breiman, 2001) is a machine-learning technique that can be used to quantify and interpret the contribution of environmental drivers (covariates) to the variability of different C fluxes (response variables) by combining multiple individual regression trees. This technique has been increasingly utilized to upscale global C fluxes from eddy covariance data (Zeng et al., 2020) but also to evaluate controls on C cycle processes (Zhang et al., 2017; López-Blanco et al., 2017, 2020). Here, we calculate the relative importance of air temperature (Ta), net radiation (Rn), soil water content (SWC), and vapour pressure deficit (VPD) controlling the C sink strength, photosynthesis, and respiration variability. This random forest algorithm constructs multiple (1000 in this analysis) decision trees during training time with different random subsamples (with replacement) from the same input training dataset. In each cluster classified by random forest, the algorithm generates a multiple linear regression to characterize different C fluxes as a function of environmental drivers (López-Blanco et al., 2017, 2020). This algorithm version (Pedregosa et al., 2011) estimates the relative importance of each covariate between 0 % and 100 %, which correspond to the fraction of decision participating during data clustering. We used the random forest algorithm to evaluate the diurnal and seasonal patterns of the relative importance of Ta, Rn, SWC, and VPD responsible for the variability of C fluxes. We used data from the June–September period aggregated per hour, and we ran multiple random forests with growing season data binned per hour of the day, day of the year, and yearly, respectively (Table S1 in the Supplement).

Since C fluxes are affected by plant phenology and climate factors, including temperature, soil moisture, and radiation simultaneously, to analyse the effect of a single factor, ideally, other factors need to be identical or at least close (no significant differences). In each comparison, data of the same period in each year were selected to exclude the influence of plant phenology. To further analyse the effect of soil moisture, radiation, and temperature on C fluxes, we selected a specific group of data for further evaluation other than the entire observation time. The group of data contains two late growing season periods: periods with a significant difference in SWC but almost identical Ta and Rn (i.e. late growing season of 2014 vs. 2015) and periods with a significant difference in Ta but almost identical SWC and Rn (i.e. late growing season of 2014 vs. 2018). Additionally, in order to analyse the effect of annual temperature on C fluxes, we selected a group of time stamps with significant differences in Ta but almost identical SWC and Rn (i.e. 2017 vs. 2014 and 2018 vs. 2014). The magnitude of the differences between C fluxes in the same group was analysed by the independent-sample t test method.

Mean daily meteorological variables (including Ta, P, SWC, and Rn) exhibited evident seasonal variability except for VPD; these variables increased progressively in the early growing seasons, reached their maximum in July, and decreased gradually afterwards (Fig. 2a, b, c, d). Air temperature during the growing season was 7.7 ± 2.6, 7.4 ± 2.6, 8.5 ± 2.9, and 9.2 ± 3.3∘C in 2014, 2015, 2017, and 2018, respectively, while precipitation totalled 662.8, 521.4, 661.2, and 624.3 mm for the same years, falling primarily during the growing season. The precipitation measured during the late growing season of 2015 was only half of the amount measured in 2014, 2017, and 2018. The lower precipitation regime led to a marked decline in SWC, making the late growing season of 2015 the driest period among all growing seasons during the study period. The greatest difference in SWC occurred in the late growing season of 2014 and 2015, when Ta values at the same period were very close (Ta was 6.8 ± 2.6 and 6.8 ± 2.5 ∘C in 2014 and 2015, respectively). Compared to 2014, SWC decreased by 15.4% in the late growing season of 2015 (Fig. 2b; Table S2). Meanwhile, SWC in the late growing season of 2014 and 2018 was almost identical (80.7 ± 4.1 and 80.8 ± 3.8, respectively), but the temperature difference was the largest (25 %) compared to any other years (Fig. 2a; Table S2).

During the growing season, NEE (Fig. 3a1–d1) and GPP (Fig. 3a3–d3) featured a clear peak of the diurnal variations; both fluxes reached their summit between 12:00 and 14:00 local time. Re, however, presented a lower daily variability. The rates of NEE, Re, and GPP during the growing season averaged -2.3 ± 0.3, 3.2 ± 1.0, and 5.5 ± 0.9 µmol m-2 s-1, respectively, for the entire study period (2014, 2015, 2017, and 2018). For the late growing seasons, the lowest rate of net C uptake was measured in 2015 (-10.0 µmol m-2 s-1), whereas 2014 (-12.4 µmol m-2 s-1), 2017 (-12.2 µmol m-2 s-1), and 2018 (-12.5 µmol m-2 s-1) exhibited more negative NEE values (i.e. stronger net C uptake rate). Between the late growing season in 2015 and the late growing seasons in 2014, 2017, and 2018, there was a significant difference in the rates of Re (p<0.01), while no significant difference was found in GPP variability (p>0.05), suggesting that Re may be the component causing the difference observed in NEE. Specifically, the rates of Re in the late growing season of 2014 and 2015 were 2.4 ± 0.2 and 3.0 ± 0.2 µmol m-2 s-1, respectively, which indicated that the drier conditions in 2015 generated a 25 % higher Re compared to 2014. In addition, the rate of Re in the late growing season was significantly higher in warmer 2018 (3.5 ± 0.2 µmol m-2 s-1) than in 2014 (2.4 ± 0.2 µmol m-2 s-1).

NEE, Re, and GPP also exhibited a strong seasonal variability; the C fluxes gradually increased from low values in early June to the maximum in the middle of the growing season (late July to early August on average), followed by a decrease towards the end of the growing season (Fig. 4a, b, c). The C sink strength across the growing season was found lowest in 2018, followed by 2017, whereas 2014 and 2015 exhibited relatively higher values (Fig. 4a). Notably, the accumulated Re in the late growing season was significantly higher in 2015 compared to 2014 (P<0.05), while there was no significant difference in GPP (Fig. 4b; Table S2). Moreover, the late growing season of 2015 witnessed the lowest SWC while keeping the same Ta compared to the same period in 2014 (Fig. 2b; Table S2). The substantial decline of SWC observed in 2015 appeared to be responsible for the weaker observed C sink strength. On the other hand, Ta in the late growing season of 2018 was the highest for the same period among all the years, while the SWC remained the same as 2014 (Fig. 2a). The late growing season of 2018 showed overall higher GPP and Re but lower net C uptake than 2014. Significantly higher Re in 2018 caused by higher temperatures eventually led to a decrease in the C sink capacity (Fig. 4a, b, c; Table S2).

Our data processed by a machine-learning technique suggest that the relative importance of the primary environmental drivers (Rn, Ta, VPD, and SWC) regulating C fluxes varies on different timescales (i.e. diurnal and seasonal scales) in this swamp meadow of the QTP (Fig. 5a, b, c). The diurnal variability of NEE and GPP was mostly driven by Ta (Fig. 5a), especially in the central hours of the day between 11:00 and 15:00, while Re showed a fairly lower temperature dependence compared to NEE and GPP (Fig. 5a). SWC was relatively more important than air temperature controlling the diurnal variability of respiration (Fig. 5a). The seasonal variability shaping the terrestrial C fluxes are regulated not only by meteorological variables but also by plant phenology. To separate the role of meteorological variables from phenology, we carried out a random forest analysis every fortnight and assumed that plant phenology changed little during this time span (Fig. 5b; seasonal variability). The analyses based on random forest revealed a distinct seasonal pattern for different C fluxes from June to September, pointing to a marked contribution of net radiation over NEE and GPP (Fig. 5b). Re, however, was found mostly regulated by SWC. The contribution patterns of each environmental driver to the variations of C fluxes on an inter-seasonal scale (Fig. 5c) are similar to the ones found at the seasonal scale (Fig. 5b). The inter-seasonal variability of NEE and GPP was explained clearly by Rn, while SWC revealed a stronger relative importance over Re. Overall, SWC dynamics seem to be the most important variable explaining the variability observed in the Re data, suggesting that soil moisture plays an essential role on a diurnal and seasonal basis in this cold swamp meadow ecosystem. Note also that Ta played a secondary role in regulating Re at all assessed timescales (Fig. 5a, b, c).

A previous study in alpine meadow ecosystems found that water stress may be the key limiting factor leading to a decline in the photosynthetic rate at noon with a low SWC of 6 %–21 % (Zhang et al., 2018). In this alpine swamp meadow, the soil layer maintained a relatively high SWC due to the frequent precipitation during the growing season. SWC was always greater than 70 % during the entire study period (Fig. 2b). Therefore, microbial activity and thus heterotrophic respiration were likely suppressed by the anaerobic environment due to saturated soil water condition (Chimner and Cooper, 2003; Sun et al., 2021). At this site, SWC was found to be a more important factor than temperature in controlling the variability of Re at different time aggregations (Fig. 5a, b, c). This result suggests that even under water-saturated conditions, the C dynamics of this alpine swamp meadow are still highly sensitive to changes in soil moisture and could therefore be significantly influenced by future changes in water supply (Li et al., 2015). In fact, previous studies have stressed that soil moisture will likely interact with temperature to affect Re (Han et al., 2013) and therefore modify the overall C sink strength.

To better understand the underlying mechanisms around how SWC interacts with the C fluxes in the studied alpine swamp meadow ecosystem, we selected two late growing season periods, which have significant differences in SWC but no significant difference in Ta (Fig. 6a; Table S2).

The most significant difference in C fluxes between the late growing season of 2014 and 2015 was observed in Re (p<0.05). Additionally, on both diurnal and seasonal scales, a 15.4 % decrease in SWC in the late growing season of 2015 resulted in a 25.7 % increase in Re and a 19.4 % decrease in net C uptake compared to 2014 (Fig. 6a, b; Table S2). According to Jansson and Hofmockel (2020), the intensification of anaerobic conditions due to water saturated soil can be responsible for weaker respiratory losses. This finding suggests that drier conditions in 2015 likely prevented this alpine swamp meadow from water-logged states, thereby strengthening soil respiration due to improved soil aeration (Wang et al., 2014).

Although the SWC in September 2015 was much greater than the 6 %–21 % range reported by Zhang et al. (2018), our data suggest that a 22.2 % reduction in SWC in September 2015 resulted in a 51.6 % decline in net C uptake compared to September 2014 (Fig. S1; Table S3). There is evidence from the literature that the rates of net C uptake in alpine wetlands during the growing season can be lower under drier conditions (Hao et al., 2011), indicating that this alpine swamp meadow ecosystem may be adapted to high levels of SWC (Li et al., 2015). Higher SWC may limit the diffusion of oxygen from the atmosphere to the soil, inhibiting the activity of microorganisms and reducing the decomposition rate of soil organic matter (Chimner and Cooper, 2003). Our comparisons suggest that drying can weaken the overall C sink strength in this alpine swamp meadow ecosystem. Wetlands are predicted to experience lower water tables due to permafrost degradation in the Tibetan Plateau, and, therefore, permafrost thaw-induced wetland drying could enhance the response of C emissions to climate warming (Yu et al., 2020).

The important role played by temperature controlling C exchange has been extensively found in alpine marshland across the QTP (Qi et al., 2021). For example, Zhao et al. (2010) and Zhao et al. (2005) show that Re follows the exponential variation of soil temperature. Zhu et al. (2020) also suggested that soil temperature plays the most important role in the change of monthly Re in the alpine wetland at Luanhaizi, northeastern Qinghai–Tibetan Plateau. We therefore explored another comparison between the late growing season of 2014 and 2018 (Fig. 7a, b; Table S2) when phenology, radiation, and SWC were almost identical at both periods but temperature differed. Compared to 2014, a 25 % increase in Ta in the late growing season of 2018 led to joint larger GPP and Re fluxes (Fig. 7a, b; Table S2). Although both GPP and Re increased, the intensification in Re was greater than GPP, indicating that warmer temperatures had a stronger impact on Re, resulting in a decrease in the net C uptake (Fig. 7b; Table S2). To evaluate if this finding is also consistent at an annual scale, we further analysed annual aggregated data. An annual comparison was made between 2014, 2017, and 2018 when SWC was found insignificantly different while temperatures in 2017 and 2018 were 44.4 % higher than in 2014 (Table S4). Again, this 44.4 % increase in Ta in 2017 and 2018 both led to stronger GPP and Re (Table S4). Although both GPP and Re increased, the intensity in Re was greater than GPP, indicating that warmer temperatures have a stronger impact on Re at this site, resulting in an approximately 50 % decrease in the net C uptake (Table S4).

This comparison suggests that future warming could weaken the overall C sink strength in this alpine swamp meadow ecosystem. Similar climate sensitivities have also been found in recent studies. For example, a study performed by Niu et al. (2017) in an alpine swamp meadow on the central Tibetan Plateau suggests Re was more sensitive to increased temperature than GPP. This suggests that global warming may exacerbate future C releases in the alpine wetlands of the QTP (Gao et al., 2019; Niu et al., 2017; Zhu et al., 2020). Liu et al. (2018) concluded that warming has a significant inhibitory effect on GPP and a minor effect on Re, resulting in a weaker carbon sequestration capacity of their studied alpine wetland ecosystem. However, other researchers have also reached different conclusions. For instance, Qi et al. (2021) found that GPP is consistently more sensitive than Re to changes of temperature at daily, seasonal, and annual scales, suggesting that cold condition can act as a strong constraint on C uptake in alpine marshlands. Wei et al. (2021) also found that the uptake of C by plants will exceed the amount of C release under warmer and wetter climate conditions based on manipulative experiments and model simulations for the Tibetan Plateau. Their study is based on a longer-term trend while our study only covers 4 years of year-round observations; thus, site-specific differences in time and space scales may explain this variability. These inconsistent ecosystem responses suggest that there are still large uncertainties regulating the responses of C fluxes to temperature variation, and further work is still crucial.

The QTP experienced a higher rate of temperature increase than that of the Northern Hemisphere average (Zhang et al., 2013). The effects triggered by climate-induced warming over NEE in this area have been argued to either increase or decrease the net C balance NEE or even have no effect whatsoever (Ganjurjav et al., 2018; Li et al., 2020; Wu et al., 2011; Zhu et al., 2017). These inconsistent responses could be due to water limitations offsetting the C balance or even reversing the effect of elevated temperatures, which change the decomposition and photosynthetic processes (Wu et al., 2011; Yu et al., 2013; J. Zhao et al., 2019). Alpine swamp meadows of the QTP have recently attracted much attention because they hold 5.9 % (∼ 1.98 Pg C) of the total grassland soil organic C (∼ 33.52 Pg C). Such ecosystems have the highest organic C density (∼ 50 kg C m-2) and play an important role in the global C cycle (Niu et al., 2017). To test whether the observed SWC effects in this study were representative of other sites on the QTP and put it into a broader perspective, we examined the temperature and precipitation (as a proxy for SWC) impacts on NEE (Liu et al., 2016).

The NEE observations from this study were within the NEE ranges of previous studies in similar ecosystems located across the QTP (-255.5 to 173.2 g C m-2 yr-1) (Table 2). According to Wei et al. (2021), there are six observational studies about C flux around our study site, and three of them are focused on alpine swamp meadows. Among them, one study had a 1-year dataset (Zhang et al., 2008), and the other two characterized the same location (Zhao et al., 2005, 2010). The three studies were reported as a net C source, while our 4-year dataset revealed that this alpine swamp meadow functioned as a net C sink of -168.0 ± 62.5 g C m-2 yr-1 at a 3571 m a.s.l. The different directions of C exchange suggest that there are still uncertainties in our understanding of C exchange in alpine swamp meadows, and further efforts are still needed to improve our projection of C balance change of this ecosystem under changing climate.

In addition, the NEE estimates of this alpine swamp meadow show a stronger C sink strength than those from alpine meadows (-161.3 to 85.4 g C m-2 yr-1) (Chai et al., 2017; Wang et al., 2017; Wu et al., 2020), alpine steppes (-30 to 21.8 g C m-2 yr-1) (Wang et al., 2018; D. Wang et al., 2020; Y. Wang et al., 2020; Wu et al., 2010), and alpine shrublands (-14 to -67 g C m-2 yr-1) (Zhao et al., 2005, 2006). This is likely a result of the inhibiting effects of the nearly saturated soil condition over soil respiration (Sun et al., 2021). In permanently or seasonally inundated swamp meadows, high SWC may have triggered lower C loss rates and further benefited C preservation. At our site, the higher C sink strength was likely attributed to higher precipitation (and therefore higher SWC) and lower temperature, which created colder and more humid conditions than other sites (Table 2). It has been demonstrated that cold and humid conditions favour stronger C sinks in alpine meadow ecosystems (Fu et al., 2009).

The interannual comparison of the sites presented in Table 2 shows that under low annual precipitation conditions (∼ 300 mm), the joint effects of warming and reduced precipitation weakened the net C uptake at Damxung (Niu et al., 2017) and even turned the C sink of Xiaobo Lake wetland into a C source when comparing 2015 with 2012 and 2013 (Cao et al., 2017; Wu et al., 2018). Under relatively high annual precipitation (∼ 500 mm), the joint effects of warming and increased precipitation enhanced C release in the Haibeia site (see Table 2) when comparing 2006 to 2004 and 2005 (Zhao et al., 2010). This indicates that net C uptake under warming conditions can be weakened even under high annual precipitation rates.