Influences of drought on the stability of an alpine meadow ecosystem

ABSTRACT Drought plays a prominent role in affecting ecosystem stability and ecosystem productivity. Based on eddy covariance and climatic observations during 2012–2020, the Fisher discriminant analysis method was employed to accurately detect drought occurrences. Furthermore, the ecosystem water sensitivity and its resistance to drought were quantified to evaluate the ecosystem stability. The results showed that the alpine meadow suffered drought most frequently at the beginning of the growing seasons. However, drought during the peak growing seasons reduced the gross primary productivity (GPP) the most, by 30.5 ± 15.2%. In the middle of the peak growing seasons, the ecosystem water sensitivity was weak, and thus, the resistance to drought was strong, which resulted in high ecosystem stability. At the beginning and end of the peak growing seasons, the ecosystem stability was relatively weak. Ecosystem stability was positively related to the corresponding multiyear average soil water content (SWCave). However, drought occurring during high SWCave periods led to larger reductions in GPP, which indicated that the inhibitory effects of drought on ecosystems were more dependent on the occurrence time of droughts than on ecosystem stability.


Introduction
Under global climate change, drought events occur more frequently and intensively , which greatly affects the stability of the ecosystem and hence the global carbon budgets (Sheffield, Wood, and Roderick 2012). In the meteorological research field, droughts induced by environmental factors are defined as soil drought and atmospheric drought, which are caused by low soil water content (SWC) and high atmospheric water vapor pressure deficit (VPD), respectively. The water in soil is the source for plant transpiration and soil evaporation and greatly affects material and energy cycling (Heathman et al. 2012;Sheffield, Wood, and Roderick 2012;Ma et al. 2019). Therefore, soil drought is considered the main form of dryness stress (Liu et al. 2020a), and SWC is an important index for soil drought discrimination (Beier et al. 2012;Vanderlinden et al. 2012). In addition to soil drought, atmospheric drought has received increasing attention due to its critical role in determining plant functioning (Novick et al. 2016).
In the past 30 years, more frequent drought events have been observed and recorded on the Qinghai-Tibet Plateau, which have inhibited meadow growth more seriously (Ding et al. 2018). Meanwhile, at the Naqu station, Xu et al. (2021) determined soil and atmospheric drought events by SWC and VPD thresholds, respectively. The results showed significant reductions in gross primary productivity (GPP) induced by both soil and atmospheric droughts. Moreover, YuYuan et al. (2019) suggested that the inhibiting effect of atmospheric drought on vegetation growth was even stronger than that of soil drought. Therefore, to further study the effects of drought on ecosystem stability, it is necessary to consider both soil and atmospheric water conditions to accurately determine drought events.
However, few studies provide an effective drought detection method that comprehensively takes soil and atmospheric water conditions into consideration. Therefore, in this study, we employed the Fisher discriminant analysis method to detect the occurrence of drought in the alpine meadow of northern Tibet. The Fisher discriminant method uses the variance analysis method, follows the principle of extreme value of differentiation, makes use of known sample categories to maximize the ratio of interclass variance to intraclass variance, and establishes a discriminant function as a classifier to judge unknown data categories. This method is widely used to solve taxonomic problems (Pan 2013) and provides us with an effective method to solve drought detection problems. Fisher discriminant analysis could be applied when the events to be classified have multiple variable features. Therefore, the soil moisture and atmospheric moisture conditions can be comprehensively considered in drought identification. After the accurate detection of drought events, the effects of drought on ecosystem stability can be further studied in detail.
Grassland ecosystems are mainly constructed of herbaceous plants with shallow root systems. Due to the short life spans of grass, the material and energy cycles are relatively fast in grassland ecosystems. Therefore, grasslands are usually more vulnerable to extreme climatic events, especially drought (Cook et al. 2014;Novick et al. 2004). Previous studies have indicated that the productivities of grassland ecosystems are sensitive to water availability and are even dominated by water conditions (Wolf et al. 2013). At the beginning of drought events, grass can absorb water in the shallow soil layer to support normal photosynthesis and transpiration, and thus, there are no instant decreases in productivity (Teuling et al. 2010;Wolf et al. 2013). However, the shallow soil water would be depleted as the drought continued. Consequently, the aboveground biomass and vegetation coverage would decrease, which would lead to rapid decreases in ecosystem productivity and stability (Huxman et al. 2004;Ponce-Campos et al. 2013;Hoover, Knapp, and Smith 2014). Accordingly, vegetation productivity could be used as an effective indicator to evaluate ecosystem sensitivity and resistance to drought due to its sensitivity to drought (Knapp et al. 2015;Stuart-Haentjens et al. 2018).
Known as the "third pole of the earth" and the "water tower of Asia," the Qinghai-Tibet Plateau has an extremely important ecological strategic position. The widely distributed alpine meadow ecosystem on the Qinghai-Tibet Plateau plays an important role in maintaining the ecological stability of the plateau. Meanwhile, as alpine meadows are fragile and sensitive to disturbances, studies in this area would be indicative of global climate change. As previous studies indicated the determinative role of the water condition on the productivities of alpine meadow ecosystems (Fu, Shen, and Zhang 2018;Zhang et al. 2018), drought would probably be the most concerning climatic event that greatly threatens the stability of alpine meadow ecosystems (Xu et al. 2021). Research on the effects of drought on ecosystem stability on the Qinghai-Tibet Plateau would be beneficial for protecting the ecological barrier function of Tibet and coping with global climate change.
Previous studies on ecosystem stability were conducted either on an annual scale by collecting productivity data from different sites (Knapp et al. 2015;Stuart-Haentjens et al. 2018) or on a relatively short time scale by control experiments (Zhang et al. 2017). Although the variabilities in ecosystem stability under extreme water conditions have been studied (Hoover, Knapp, and Smith 2014), there is still a knowledge gap regarding the specific effects of drought on ecosystem stability in different stages during the growing seasons (Evans et al. 2011). Therefore, this study will take advantage of the continuous flux data observed by the eddy covariance technique and hence directly evaluate the impact of drought disturbance on ecosystem stability in different periods of the growing season, which could capture the true stability changes of natural ecosystems.
This study was conducted in a typical alpine meadow ecosystem at Naqu station in northern Tibet. Fisher discriminant analysis was originally applied to comprehensively determine the occurrence (including onset and end time) of drought events by using SWC to represent soil water conditions and VPD to represent atmospheric water conditions. Based on the accurate drought detection results and the nine-year in situ flux observations, the effects of drought on the stability of the alpine meadow were quantified when drought events occurred in different periods during the growing seasons. The results will be helpful for understanding the mechanisms of ecosystem stability and further protecting the Qinghai-Tibet Plateau.

Site description
The study was conducted at Naqu station (31.64°N, 92.01°E, 4598 m a.s.l.), which is a member of the Chinese Flux Observation and Research Network (ChinaFLUX). The site is in northern Tibet and is located between the Tanggula Mountains, the Nianqing Tanggula Mountains and the Gangdice Mountains. This region is dominated by a plateau subfrigid monsoon climate, with a cold-dry and anoxic environment. The radiation is sufficient at this site, and the cumulative sunshine hours exceed 2886 h. The annual average air temperature is -1.9°C, and the annual precipitation is 430 mm. There is no absolute frost-free period throughout the year. The growing season generally began in June and ended in September. Good water conditions could lead to slightly earlier vegetation green-up. July to August is the peak growing season when the vegetation is relatively flourishing. The vegetation is a typical Tibetan alpine meadow, which is quite sensitive and fragile, dominated by Kobresia pygmaea, accompanied by Potentilla saundersiana, Potentilla bifurca, Potentilla cuneata, Saussurea stolickai, etc. The soil is typical alpine meadow soil with sandy loam.

Observation and instrumentation
In 2011, an open-path eddy covariance (OPEC) system was installed at the Naqu site. Since then, the carbon flux and water flux have been observed continually. The instruments were installed on an iron tower at a height of 2.3 m, which included an infrared gas analyzer (Model LI-7500A, Li-cor Inc., Lincoln, Nebraska, USA) and a 3-D sonic anemometer (Model CSAT-3, Campbell Scientific Inc., Logan, Utah, USA). The CO 2 concentration and 3-D wind speed were sampled at a frequency of 10 Hz and recorded by a CR1000 datalogger (Campbell Scientific Inc.), and the raw eddy covariances between CO 2 concentration and vertical wind speed were calculated and recorded at 30 min intervals by the CR1000 datalogger.
The meteorological factors were observed by a microclimate observing system. The net radiation (Rn) and photosynthetic active radiation (PAR) were measured by a four-component net radiometer (Model CNR-1, Kipp&Zonen, Delft, Netherlands) and a PAR sensor (LI-190SB, Li-cor Inc.) at a height of 1.5 m, respectively. The air temperature (Ta) and relative humidity (RH) were measured by a temperature and humidity probe (Model HMP45C, Vaisala Inc., Helsinki, Finland) installed at a height of 1.8 m in a radiation shield (Model 41,002, RM Young Inc.), and then the vapor pressure deficit (VPD) was calculated. A rain gauge (TE525MM-L, Campbell Scientific Inc.) was used to record precipitation (PPT) at 0.5 m height. Thermometers (109-L, Campbell Scientific Inc.) and time-domain reflectometer (TDR) probes (Model CS616-L, Campbell Scientific Inc.) were used to measure the soil temperature (Ts) and soil water content (SWC) at depths of 5 cm, 10 cm, 20 cm, and 50 cm. The data were sampled at 1 Hz and then resampled to 30 min averages by a CR1000 datalogger (Campbell Scientific Inc.).

Data processing
The data were processed according to the ChinaFLUX standard data processing methods (Yu et al. 2006), which mainly include double coordinate rotation (Aubinet et al. 2000) and the Webb, Pearman and Leuning density correction (WPL correction, Webb, Pearman, and Leuning 1980) and fake flux removal. In addition, the stable atmospheric stratification at night would cause an underestimation of carbon fluxes. Therefore, the flux data with a frictional wind speed (u*) less than 0.14 m s −1 were removed from the data set. Approximately 24.6% of the half-hour data were excluded for the daytime, and 86.3% were excluded for the nighttime. The daytime data gaps were filled using the Michaelis-Menten equation (Michaelis and Menten 1913) during the growing seasons. The Lloyd-Taylor equation (Lloyd and Taylor 1994) was used to fill the data gaps in the nighttime and nongrowing seasons.
The carbon flux measured by the OPEC system was the net ecosystem CO 2 exchange (NEE). The NEE observed during nighttime equaled the nighttime ecosystem respiration (RE), which could be extrapolated to daytime using the Lloyd-Taylor equation to obtain the whole RE data set. Then, the gross primary productivity (GPP) of the ecosystem can be obtained by equation (1).
For meteorological data, the threshold method was used to remove spurious data. The data gaps were filled using linear interpolation or the mean diurnal variation (MDV) method (Falge et al. 2001).

Fisher discriminant analysis
The Fisher discriminant analysis method is one of the commonly used discriminant methods. The basic principle of the method is to construct a linear function yc consisting of p variables (the two variables selected in this study were SWC and VPD). Since different variables correspond to different function values, a suitable yc value can be obtained as a criterion to distinguish between different objects (Pan 2013).
To construct the drought discrimination index, SWC and VPD were used as discriminant factors in this study. The SWC observed at a depth of 5 cm was selected to characterize the soil water condition, and the VPD was selected to represent the atmospheric water condition.
where VPD is the vapor pressure deficit (kPa), Ta is the air temperature (°C), and RH is the relative humidity (%). The drought days in the middle of drought events in typical drought years were selected to ensure that SWC and VPD could fully express the characteristics of drought and thereby the drought sample series could be constructed. Meanwhile, the SWC and VPD in the periods with abundant PPT were used to construct the nondrought sample series. The training sample series (Table S1) was used to calculate the sample averages of each type (m i ): where n is i = 1,2 the sample size of the class w i , and X is the training sample series.
The intraclass dispersion matrix (S i ), the total intraclass dispersion matrix (S w ), and the interclass dispersion matrix (S b ) of the sample were calculated by the following equations: Then, the optimal projection vector (w*) was calculated: When the ratio of the interclass variance to the intraclass variance reached the maximum after the samples were projected in this projection space, the initial equation of the critical line between drought and nondrought samples could be obtained: where c 1 and c 2 are the determination coefficients, x 1 is the SWC series, x 2 is the VPD series, and y c is the criterion.
As the sample size of the drought series equals that of the nondrought series, y c can be expressed as follows: where x 1D is the SWC average of drought samples, x 2D is the VPD average of drought samples, x 1N is the SWC average of nondrought samples, and x 2N is the VPD average of nondrought samples. Integrating Equations (8) and (9), the drought occurrence in the alpine meadow ecosystem could be detected according to the critical line: The meteorological data (SWC and VPD) to be discriminated could be classified according to Equation (10): y = 1 is the boundary value; when y > 1, the center of the samples was above the critical line, and drought would occur when it continued for more than 10 days (Zhang et al. 2022); when y < 1, the center of the samples was below the critical line, and drought would not occur.

Calculation of water sensitivity of the ecosystem
Based on the drought detection results of Fisher discriminant analysis in 2.4, the onset and ending dates of drought events could be determined. Variations in GPP were used to quantify the changes in ecosystem function under drought conditions (Hoover, Knapp, and Smith 2014). To match the GPP data, the SWC was used to characterize the variations in water conditions. Thus, the water sensitivity (S) of the ecosystem can be expressed in terms of ΔGPP/ΔSWC (Zhou et al. 2006). To eliminate the effects of the dimensions and units, the relative variations in GPP and SWC were calculated (ΔGPP=GPP and ΔSWC=SWC): where ΔGPP and ΔSWC are the decreases in GPP and SWC during droughts, respectively, and GPP and SWC are the averages of GPP and SWC during drought, respectively. The smaller value of S indicated that the GPP was not sensitive to water and that the ecosystem stability was stronger, and vice versa.

Calculation of ecosystem resistance
The decrease in GPP during drought was used as a biological index to measure ecosystem resistance (Zhang and Zhao 2010). During the drought period detected by the Fisher discriminant method, the ratio of the GPP after drought stress to the GPP before drought stress was used to characterize ecosystem resistance (R; Stuart-Haentjens et al. 2018): where GPP* is the GPP value after drought stress and GPP is the GPP value before drought stress. A larger R indicated stronger ecosystem resistance to drought and vice versa. To exclude the influences of drought duration on the R value, the R value was calculated on the 12 th day (12 days was the shortest duration of a drought event observed during the study period) after the onset of the drought events.

The variations in meteorological factors and GPP
Radiation resources were relatively abundant on the Tibetan Plateau with small interannual variability ( Figure 1A). Throughout the growing season, the Rn was relatively high, with values above 10 MJ m −2 d −1 (Figure 2A). The seasonal variations in Ta and Ts followed a unimodal pattern, respectively, which peaked in July or August. The interannual variabilities in Ta and Ts were also small ( Figure 1B). The PPT was mainly in the growing seasons with great interannual variability ( Figure. 1D and Figure 2C). Consequently, a shortage of PPT led to a decrease in SWC, and large decreases could be obviously observed during the peak growing seasons of 2013,2015,2016,2017, and 2019 ( Figure 2C). Droughts would probably have occurred at this time combined with the high Rn and Ta. VPD is under the control of both Ta and water conditions. It generally peaked in June ( Figure 1E). The daily values of VPD varied by approximately 0.37 kPa in the growing season ( Figure 2D). The general variation pattern of GPP followed a unimodal curve in the growing season, but it was greatly affected by water conditions and varied greatly in different years ( Figure 1F). GPP decreased sharply when drought occurred during the peak growing seasons, which sometimes broke the unimodal pattern ( Figure 2D).

The occurrence regulation of droughts and their effects on GPP
To accurately detect the occurrence time of drought, the Fisher discriminant analysis method was used in this study. Based on the 9-year observation data, the typical drought and nondrought days were used as the training samples (Table S1); hence, the discriminant equation was established as y = 24.46 SWC -4.60 VPD (Figure 3), which comprehensively took the soil and atmospheric water conditions into consideration. Accordingly, the drought events were identified and labeled with gray shadows in Figure 2.
From the beginning to the end of the growing seasons, the frequencies of droughts decreased ( Figure 4A). In mid-to-late May (day of year (DOY) 131-149), droughts occurred most frequently. The frequency was as high as 50%. In two periods of the peak growing seasons (DOY 220-223 and DOY 231-238), the frequency of droughts reached 33%. According to the 9-year observations, there were no drought events in early July and September , which might be a drought-free period in this alpine meadow.
During the study period, the alpine ecosystem was only nearly drought free in the growing seasons in 2020. The GPP showed a typical unimodal variation in 2020, with the highest annual cumulative value (Table  S2). Drought was only found at the end of the growing season in 2020 and had little influence on GPP ( Figure 2). Therefore, the GPP of 2020 was selected as the baseline for evaluating the GPP decrease rate induced by droughts that occurred in other years. Accordingly, the inhibiting effects of droughts occurring in different periods during the growing seasons could be quantified ( Figure 4B). The results indicated that drought during the peak growing seasons decreased GPP by 30.5 ± 15.2%; drought at the beginning and end of the growing seasons decreased GPP by 17.1 ± 42.1% and 12.4 ± 11.4%, respectively. The inhibiting effects of drought at the beginning of the growing seasons had great uncertainty, which might be due to the great interannual variabilities in the start time of the real growing season.

The effects of drought on ecosystem stability
When drought occurred, vegetation growth was inhibited due to the shortage of soil water supply, high atmospheric water demand, strong radiation and high temperature (Figure 2). GPP did not decrease obviously when drought occurred at the beginning of the growing seasons. However, the drought events that occurred during the peak growing seasons led to great decreases in GPP ( Figure 2D) and thus great losses of annual cumulative GPP ( Figure 4B). Therefore, we focused on the effects of drought disturbances on alpine meadow ecosystem stability in peak growing seasons (July-August).
Drought occurred quite frequently in the alpine meadow ecosystem. During the study period (2012-2020), drought occurred in the peak growing seasons of 2013, 2015, 2016, 2017 and 2019, which were five of the observed nine years (Table 1). The year 2015 was special; it was extremely droughty, with an annual PPT as low as 289.7 mm (Table. S2). During drought in the peak growing season of 2015, the SWC was close to that in 2016, 2017 and 2019, but the GPP was lower than that in these years. At the beginning of the growing season in 2015, the PPT was only 1.4 mm with high VPD, and the GPP was greatly depressed. In summary, in the extreme drought year of 2015, the alpine meadow ecosystem presented relatively high water sensitivity and low resistance stability. To exclude the disturbance of this extreme sample, the years 2013, 2016, 2017 and 2019 were selected to further evaluate the stability of this alpine meadow in peak growing seasons.
During the peak growing seasons, the variations in the water sensitivity and the resistance to drought of this alpine meadow ecosystem could be modeled by quadratic functions (Figure 5A and Figure 5B). In the middle of the peak growing seasons (from late July to early August, DOY 205-215), the ecosystem water sensitivity was lowest, and the resistance was strongest, which indicated that the stability of this ecosystem was  highest in this period. At the beginning and end of the peak growing seasons, the ecosystem water sensitivity obviously increased, and the resistance decreased, which led to a decrease in the ecosystem stability. Throughout the whole process, ecosystem resistance was found to be negatively related to water sensitivity ( Figure 5C).
The water sensitivity and resistance of the alpine meadow ecosystem in the peak growing seasons were significantly regulated by the corresponding multiyear average SWC (SWC ave ; Figure 6). The water sensitivity decreased exponentially with increasing SWC ave ( Figure 6A), and the resistance increased logarithmically with increasing SWC ave ( Figure 6B). The water sensitivity (resistance) of the ecosystem was not found to be significantly regulated by other meteorological factors.
When the ecosystem water sensitivity was low and the ecosystem resistance was strong , the water condition was generally good (Figure 6), and the vegetation grew well. If drought occurred at this time, the decrease in GPP would be larger than in wet years (Figure 7). When the ecosystem water sensitivity was high and the ecosystem resistance was weak (DOY230-242), the water conditions were usually poor (Figure 6), and the vegetation productivity was relatively low. If drought occurred at this time, the reduction in GPP would be relatively small (Figure 7). Therefore, the high ecosystem stability did not absolutely mean small effects of drought on GPP. The inhibiting effects of drought on ecosystem productivity may depend more on the timing of drought.

The effects of drought disturbances on the carbon flux of the alpine meadow ecosystem
The water condition plays a critical role in determining the ecosystem productivity in this alpine meadow ). In addition, water stress occurred quite frequently during its relatively short growing seasons (Figure 2). The droughts at different periods during the growing seasons (at the beginning of the growing season, at the peak growing season, or at the end of the growing season) had divergent effects on the GPP of the ecosystem. Moreover, soil drought and atmospheric drought have been reported to have different effects (Xu et al. 2021).
In the study area, the climate is arid, and 80% of the PPT falls in the growing seasons . The spring drought easily occurred due to the small PPT in winter. Drought also occurred frequently during the growing seasons, although the PPT was greater than that in the nongrowing seasons. A regulation could be found about the drought frequencies during the growing seasons. It occurred most often at the Figure 4. The frequencies of droughts during the growing seasons in the alpine meadow ecosystem (A). The decrease rate of GPP induced by droughts occurred during different periods in the growing seasons (B). In this figure, as long as the discriminant y > 1, drought was considered to have occurred on that day. Table 1. The characteristics of climatic factors, the water sensitivity and drought resistance of the alpine meadow during droughts. DOY-Day of year; Ta-Air temperature; Ts-Soil temperature; PPT-Precipitation; SWC-Soil water content; VPD-Vapor pressure deficit; GPP-Gross primary productivity. beginning of the growing seasons, followed by the peak growing seasons. Drought was found to have the lowest frequency at the end of the growing seasons (Figure 4). In the peak growing seasons, the vegetation grew well, and more water was transferred to the atmosphere through transpiration . Meanwhile, the evaporation would be larger due to the high Ta. Under the effects of stronger transpiration and evaporation, there would be more water vapor in the atmosphere. Consequently, soil drought occurred more frequently than atmospheric drought during this period (Xu et al. 2021). The soil water directly provided water for vegetation growth. Therefore, the GPP decreased sharply when drought occurred during the peak growing seasons due to soil water depletion (Figure 2; Granier et al. 2007). Drought in the peak growing seasons led to greater annual cumulative GPP loss than drought at the beginning or end of the growing seasons ( Figure 4B). Atmospheric drought occurred more frequently at the beginning of the growing seasons (Xu et al. 2021). At this time, the GPP was low, and thus, the inhibitory effect of drought on GPP was relatively small (Ma et al. 2019). However, water stress postponed the start time of the growing seasons (Luo et al., 2021), which further affected the annual cumulative GPP (Zhang et al. 2011).   Consequently, the occurrence time and characteristics of droughts would have different effects on carbon fluxes, which showed great uncertainties according to previous studies (Novick et al. 2004;Xu et al. 2021). Therefore, it would be profoundly meaningful to comprehensively detect droughts according to both soil and atmospheric water conditions (Liu et al. 2020a;YuYuan et al. 2019), to quantify the ecosystem water sensitivity and resistance to drought and to quantitatively assess the effects of drought. This would greatly improve the understanding of this alpine meadow ecosystem and help predict its carbon sink or source properties under global climate change.

The resistance stability of the alpine meadow ecosystem to drought
The alpine meadow ecosystem is fragile and sensitive to climate change. Its resistance stability under drought conditions determines its adaptation to global climate change (Hoover, Knapp, and Smith 2014). In this study, the water sensitivity and drought resistance of the ecosystem were used to evaluate the resistance stability of the alpine meadow under drought disturbances. At the regional scale, the water sensitivity of the ecosystem decreased exponentially with the multiyear average PPT (Huxman et al. 2004;Knapp et al. 2015), whereas the resistance increased linearly with the multiyear average PPT (Stuart-Haentjens et al. 2018). Similar regulation was found during the peak growing seasons in this study. The difference was that the relations between the ecosystem water sensitivity/resistance were significantly related to SWC ave instead of multiyear average PPT ( Figure 6). In the wet seasons, lower water sensitivity and higher resistance were observed in this alpine ecosystem. However, in the relatively dry seasons, the water sensitivity increased, and the resistance to drought decreased. In this study, the water sensitivity and resistance were calculated, and their relations with SWC ave were quantified in the peak growing seasons ( Figure 6). Accordingly, we could infer that the resistance stability to drought should be weaker at the beginning or end of the growing seasons than that in the peak growing seasons due to the lower SWC ave (Figure 2). Hoover, Knapp, and Smith (2014) indicated that the heatwave but not the water condition was the determinative factor for ecosystem resistance. However, the high temperature effects did not emerge in this alpine meadow ecosystem. As only natural drought events were adopted to conduct this study, the data points were limited. Although the statistical results are significant, they still have some uncertainty. Accordingly, longer observations in future studies are encouraged to further verify these results.
Previous studies indicated that in addition to environmental factors (Knapp et al. 2015;Stuart-Haentjens et al. 2018;Zhang et al. 2017), biotic factors, such as phenology, coverage, and functional diversity of the vegetation community, would probably affect ecosystem resistance (Dıáz and Cabido 2001;Evans et al. 2011).
The critical period of water is the period when the vegetation is most sensitive to the water supply. Generally, it is the transition period when the plants turn from vegetative growth to reproductive growth, which is embodied in the differentiation of flower buds. Kobresia pygmaea generally flowed in the first half of July (Xi et al. 2015); thus, the alpine meadow ecosystem was most sensitive to water and had lower resistance to drought during this time. In the fruiting period (approximately the red shadowed period in Figure 5, DOY 205-215), the ecosystem showed lower water sensitivity and higher resistance to drought. This could be attributed to the most vigorous vegetation and the highest functional diversity of the community during the middle of the peak growing season (Hooper and Vitousek 1997). The high functional diversity of the community indicated that different species were at different stages of growth with different water sensitivities, which could help the ecosystem defend against drought to some extent (Dıáz and Cabido 2001). In drought-prone ecosystems, this compensation effect between species would be more important (Hallett et al. 2014), and the dominant species played the most important role (Evans et al. 2011;Knapp et al. 2015). In addition, the ecosystem had strong resistance stability to drought disturbance during this period, which also reflected the adaptability of alpine vegetation to efficiently utilize limited heat resources to complete the growth cycle (Zhou et al. 2006). After this period, as the seeds matured, the sensitivity of the alpine meadow to water increased gradually, and its resistance to drought decreased gradually.
Drought led to greater decreases in GPP when it occurred in the period with high ecosystem resistance (Figure 7). This would be reasonable because the vegetation tended to close its stomata to maintain proper water potential to defend against dehydration, which in turn led to a reduction in photosynthesis (StPaul et al. 2012;Xu and Zhou 2008). The stronger the resistance was, the stronger these responses would be, which would lead to greater GPP decrease rates. In addition, the stronger ecosystem resistance was usually accompanied by good water conditions, at which time the vegetation coverage was high with a large plant density. Therefore, the species would compete for water resources (Evans et al. 2011). If drought occurred, the competition would be intensified, which would probably lead to decreases in the vegetation growth rate. Even worse, some of the plant individuals die due to dehydration (Hoover, Knapp, and Smith 2014). Consequently, the GPP of the ecosystem would decrease. The dominant species (Kobresia pygmaea) of this alpine meadow ecosystem had a relatively shallow root distribution and weak competitiveness for water resources . When the water competitive selection mechanism worked on the dominant species (Evans et al. 2011), it would lead to a great decrease in GPP (Hillebrand, Bennett, and Cadotte 2008).
In the extreme drought year of 2015, the ecosystem stability decreased. The ecosystem was under water stress most of the year, the growth of the vegetation was greatly depressed (Chen et al. 2020), and the annual GPP was only 62% of the multiyear average (Table S2). At the beginning of the growing seasons, the GPP increased even when droughts occurred, which postponed the increasing trend. However, in 2015, the drought that occurred at the beginning of the growing season led to a decrease in GPP. Moreover, the drought that occurred in the peak growing season of 2015 led to a greater decrease in GPP than in other years, which was probably ascribed to the severe effects of drought on ecosystem structure and function. Hence, the ecosystem failed to establish a stable homeostasis mechanism (Odum 1969;Zhang et al. 2016) and led to a decrease in ecosystem stability (Frank et al. 2015), which weakened the drought resistance of this alpine ecosystem. The other possible explanation is that the vegetation changed its life strategy to survive drought (Liu, Feng, and Fu 2020b). More photosynthate was allocated to roots to acquire more water . Consequently, the observed GPP decreased.

Conclusions
(1) Drought events occurred quite frequently in the alpine meadow ecosystem. From the beginning to the end of the growing seasons, the frequencies of drought events decreased. Drought at the beginning of the growing seasons decreased GPP by 17.1 ± 42.1%. Drought during the peak growing seasons decreased GPP by 30.5 ± 15.2%. Drought at the end of the growing seasons decreased GPP by 12.4 ± 11.4%. In the extreme drought year of 2015, the droughts throughout the year all led to sharp decreases in GPP, and the decrease amplitudes were greater.
(2) The variations in water sensitivity and drought resistance of this alpine ecosystem followed a quadric curve in the peak growing season, respectively. At the beginning and end of the peak growing season, the ecosystem had higher water sensitivity and lower resistance. In the middle of the peak growing season, the ecosystem had the weakest water sensitivity and the strongest drought resistance stability. (3) The water sensitivity and resistance of the alpine meadow ecosystem were significantly related to the SWC ave . The ecosystem stability increased with SWC ave . However, drought occurring in periods with higher ecosystem stability could also lead to greater GPP loss. The occurrence time of droughts, instead of the ecosystem stability, was more determinative to the annual cumulative GPP in this alpine meadow.