Timing and magnitude of drought impacts on carbon uptake across a grassland biome

Although drought is known to negatively impact grassland functioning, the timing and magnitude of these impacts within a growing season remain unresolved. Previous small‐scale assessments indicate grasslands may only respond to drought during narrow periods within a year; however, large‐scale assessments are now needed to uncover the general patterns and determinants of this timing. We combined remote sensing datasets of gross primary productivity and weather to assess the timing and magnitude of grassland responses to drought at 5 km2 temporal resolution across two expansive ecoregions of the western US Great Plains biome: the C4‐dominated shortgrass steppe and the C3‐dominated northern mixed prairies. Across over 700,000 pixel‐year combinations covering more than 600,000 km2, we studied how the driest years between 2003–2020 altered the daily and bi‐weekly dynamics of grassland carbon (C) uptake. Reductions to C uptake intensified into the early summer during drought and peaked in mid‐ and late June in both ecoregions. Stimulation of spring C uptake during drought was small and insufficient to compensate for losses during summer. Thus, total grassland C uptake was consistently reduced by drought across both ecoregions; however, reductions were twice as large across the more southern and warmer shortgrass steppe. Across the biome, increased summer vapor pressure deficit (VPD) was strongly linked to peak reductions in vegetation greenness during drought. Rising VPD will likely exacerbate reductions in C uptake during drought across the western US Great Plains, with these reductions greatest during the warmest months and in the warmest locations. High spatiotemporal resolution analyses of grassland response to drought over large areas provide both generalizable insights and new opportunities for basic and applied ecosystem science in these water‐limited ecoregions amid climate change.


| INTRODUC TI ON
Much of our understanding about how drought impacts the functioning of terrestrial ecosystems is based on assessments of the magnitude of the response conducted at annual timescales Kröel-Dulay et al., 2022;Luo et al., 2021;Stuart-Haëntjens et al., 2018). However, there is growing evidence that ecosystem responses to drought occur primarily during short time periods within a year, rather than over the course of the year as a whole Denton et al., 2017;Dietrich & Smith, 2016;Wolf et al., 2016). At the same time, historical reconstructions and future projections of climate change point to increased water stress and limitation that are likely to occur during the driest months of the year (Padrón et al., 2020) and within many of the driest regions on Earth (Bradford et al., 2020;Denissen et al., 2022;Huang et al., 2016). Yet, relative to our understanding of the magnitude of responses at annual timescales, our understanding of the timing of ecosystem response to drought at intra-annual timescales remains surprisingly limited.
Of the major terrestrial biomes, grasslands are among the most expansive and most sensitive to precipitation variability (Felton et al., 2021;Maurer et al., 2020;Petrie et al., 2016;Sala et al., 1988) and especially to drought (Haddad et al., 2002;Hoover et al., 2014;Lei et al., 2020;Weaver et al., 1935). Grasslands of the western US Great Plains have a strong positive correlation between monthly temperature and precipitation and a well-defined meteorological spring (March-May) and summer (June-August) (Epstein et al., 1999;Knapp et al., 2020). This period encompasses much of the growing season (Lauenroth et al., 2014) and provides context for assessing the timing of vegetation responses to dry years. Within this period, a small number of field observations and experiments indicate that grassland primary productivity is most sensitive to drought during the early-to-mid-summer months Denton et al., 2017;Dietrich & Smith, 2016;Knapp et al., 2020). Yet, it is unclear whether this temporal response pattern is consistent or varies across the biome.
Variation in the timing of grassland responses to drought may arise from different mechanisms. Figure 1 explores three different temporal responses of ecosystem carbon (C) uptake (i.e., gross primary productivity or ecosystem-level photosynthesis) throughout a growing season. While drought is hypothesized to reduce C uptake to some extent, the conceptual model considers how such impacts may vary through time. The first scenario hypothesizes that drought impacts on C uptake are consistent throughout the growing season ( Figure 1-dashed black line). In other words, drought consistently reduces C uptake by the same magnitude. This scenario assumes that temporal variation in temperature, humidity, soil moisture, phenology, or other factors do not alter the response of C uptake to drought throughout the growing season.
The second scenario hypothesizes that drought impacts are instead temporally dynamic and are greatest during the early growing season (Figure 1-dashed blue line). In other words, the impacts of drought on C uptake decrease throughout the growing season. Both biotic and abiotic mechanisms could produce this pattern. Vegetation phenology may be especially important. For example, vegetation that is most active during the cooler, early growing season (e.g., C 3 plants; Yamori et al., 2014) may be especially vulnerable to hotter and drier conditions that evolve during drought, resulting in greater reductions in early growing season C uptake.
F I G U R E 1 Conceptual model of hypothesized intra-annual responses of ecosystem carbon uptake to drought. The conceptual model assumes that drought generally decreases carbon uptake but envisions three different temporal response dynamics. The null model (black dotted line) hypothesizes that reductions in carbon uptake are consistent throughout the growing season. The red and blue dotted lines hypothesize that carbon uptake within an ecosystem should exhibit temporal variation in its response, as factors such as temperature and vegetation phenology change throughout the growing season. The red line hypothesizes that reductions in carbon uptake are minimal during the early growing season and increase into the summer, whereas the blue line hypothesizes that reductions in carbon uptake peak early on and decrease with time.
The third scenario hypothesizes that drought impacts increase throughout the growing season (Figure 1-dashed red line). The combined effects of high vegetation water use in spring and high atmospheric water demand in summer (Lian et al., 2020;Wolf et al., 2016) may exacerbate water limitation.
Increased vapor pressure deficit (VPD) is common during drought (Gamelin et al., 2022), reduces vegetation productivity (Grossiord et al., 2020;Yuan et al., 2019), and is likely to be greatest during the warmest months . Recent work in Great Plains grasslands also suggests that the proportion of cool-season precipitation increases during drought . This shift in precipitation seasonality provides a mechanism for how greater drought impacts during summer could emerge by decreasing the relative availability of water simultaneous to the highest atmospheric water demand.
The objective of this study was to determine how drought impacts to grassland C uptake vary within a growing season. We used high resolution, remotely sensed estimates of vegetation activity (e.g., gross primary productivity) and weather (e.g., precipitation) between 2003-2020 to study the timing and magnitude of grassland responses to drought across the shortgrass steppe and northern mixed prairies of the western United States. Specifically, we asked (1) When during the growing season are drought impacts to grassland C uptake greatest? (2) How is the timing of C uptake altered during drought? (3) What is the magnitude of spatial variation within and among ecoregions, and (4) Why does C uptake change during drought? By using data available on a 16-day timestep at 5 km resolution across more than 600,000 km 2 spanning nearly a 20-year period, we provide a new perspective on the spatial dynamics of the timing of drought impacts across these grassland ecoregions.

| Study region
We focused on two expansive grassland ecoregions in the western US Great Plains: the semi-arid shortgrass steppe and the northern mixed prairies (Figure 2a). These ecoregions were delineated according to Küchler's Potential Natural Vegetation map (Kuchler, 1964) and pixels were filtered to remove agricultural areas and nongrassland/shrubland land cover types based on a multi-step procedure (outlined in Felton et al., 2021). This procedure previously showed that >80% of net primary productivity in the filtered area was dominated by grasses and forbs. Overall, there were 11,307 pixels for the shortgrass steppe and 29,378 pixels for northern mixed prairies, encompassing an area of 629,484 km 2 . Mean growing season (defined broadly as March-October) precipitation ranged from 150 to 774 mm and mean growing season temperature ranged from 4.3 to 23.2°C.
Mean net primary productivity in the study area ranged from 32 to 411 g C m −2 year −1 (Felton unpublished data). Net primary productivity in the shortgrass steppe, on average the warmer ecoregion, is controlled by warm-season C 4 grasses (Lauenroth & Burke, 2008). Net primary productivity in the northern-mixed grass prairies is controlled by both cool (C 3 ) and warm (C 4 ) season grasses (Augustine et al., 2018). Sensitivity of annual net primary productivity to annual precipitation also varies considerably across these ecoregions (Felton et al., 2021).

| Vegetation carbon uptake
We acquired vegetation carbon (C) uptake data from the Moderate Resolution Imaging Spectroradiometer (MODIS) gross primary productivity data product. MODIS gross primary productivity is based on the widely accepted MOD17 algorithm, which is also the only algorithm used to produce regularly updated NASA MODIS products (Running et al., 2004). In brief, the MOD17 algorithm is based upon light-use efficiency logic established by Monteith (1972), who proposed that net primary productivity is linearly related to the amount of solar energy absorbed by the vegetation canopy. Thus, the algorithm converts (via a conversion efficiency) the absorbed photosynthetically active radiation (related to the total amount of leaf area) to C uptake. The intraannual MODIS data product describes the cumulative amount of C uptake over an 8-day period (g C m −2 8-days −1 ) across a calendar year.
We specifically acquired the MYD17A2H MODIS 8-day cumulative gross primary productivity data product (2003-2020; 18 years total); hereafter, referred to as 'C uptake.' To facilitate downstream data processing and analyses, we coarsened the resolution of the data by summing to a 16-day temporal resolution (g C m −2 16-days −1 ) and aggregating (by averaging pixel values) from 500 m to 5 km spatial resolution. We conducted follow-up analyses on a random subset of pixels (400 per ecoregion) at a finer spatial resolution (1 km, to match the native resolution of the gridded climate data used) and determined that the inferences from our analyses were not sensitive to this data coarsening. We filtered these data to represent C uptake occurring between 26 February to 24 October. This provided a purposefully broad period to encompass the full potential growing season across the two ecoregions. Indeed, an analysis using a random subset of sites (100 in each ecoregion) showed that, on average, a small fraction (<10% for both ecoregions) of total C uptake occurred in the months not analyzed ( Figure S1). We also accessed and aggregated the MODIS Normalized Difference Vegetation Index (NDVI) data product (vegetation greenness from a reflectance signal) following the same workflow to compare with inferences drawn from MOD17 gross primary productivity (a model output) and found no evidence for differences in the qualitative results (Supporting Information). Data were accessed using the MODIStools package in R, which is a programmatic interface with the MODIS Land Products Subsets offered through Oak Ridge National Laboratory. All data acquisition, processing, and analyses were conducted in R version 4.03 (R Core Team, 2022).

| Weather
We acquired temperature and precipitation data from the Daymet continuous gridded daily surface V4 weather data product (Thornton et al., 2021). The Daymet product uses statistical procedures to interpolate and extrapolate ground-based weather station measurements of daily minimum and maximum temperature, as well as daily precipitation, from weather stations onto a 1 km grid for North America. We aggregated these data to match the spatial resolution and temporal coverage of the vegetation datasets to produce 18 years of total (summed) growing season precipitation (mm) and average growing season temperature (°C), calculated as the average of the daily average of the minimum and maximum temperature.
Growing season precipitation (March-October) was highly correlated with total annual precipitation (Spearman's ρ ~.97, p < .01 in both ecoregions, Figure S2), meaning a wet or dry growing season reflected a wet or dry year. We further split these data into meteorological spring (March-May) and summer (June-August) to analyze data at seasonal scales. Data were accessed using the daymetr package in R (R Core Team, 2022).
We also acquired VPD data (kPa) from the Parameter-elevation Relationships on Independent Slopes Model (PRISM) gridded climate data product (Daly et al., 2008) for the same 18-year period. The PRISM framework also interpolates and extrapolates weather station data but specifically assumes that elevation is the dominant factor determining regional climatic variability. For these data, we focused on daily maximum VPD. We randomly selected 1.5% of pixels from the northern mixed prairies (468 pixels total) and 4% of pixels from the shortgrass steppe (471 pixels total), as stratified by latitude, to allow for a random selection of pixels homogenously spread across each ecoregion ( Figure S3). We aggregated these data from their native resolution of 4 km to 5 km to match the spatial resolution of the other datasets. Data were acquired using the PRISM Data Explorer (https://prism.orego nstate.edu/explo rer/bulk.php).

| Timing, magnitude, and progression of drought impacts
We defined 'drought' as the year with the lowest total growing season precipitation in the 18-year record for each pixel. If drought can be conceptualized as a 'prolonged absence or marked deficiency of precipitation,' then the driest year on record should be representative of this definition (Slette et al., 2019). Our approach did not focus on any one specific drought event (e.g., 2012 US drought; Wolf et al., 2016), or a particular statistical threshold of precipitation (e.g., 1st or 5th percentile; Felton et al., 2019), but instead on the driest year for each pixel. Thus, we employed a meteorological definition of drought (Ukkola et al., 2020), but acknowledge that drought is a complex phenomenon and can be defined in many ways, ranging from ecological (Munson et al., 2021) to socioeconomic drought (Mehran et al., 2015). All our analyses compared estimates from the single driest year for each pixel to all other years of data in that pixel (excluding the driest year, 17 years).
To produce a time series of C uptake for each pixel for each year (one time series per pixel per year), we used the 16-day C uptake data. From these data, we also calculated the cumulative amount of C uptake that occurred by each day of the growing season. This resulted in 203,526 pixel-year time series of 16-day and cumulative C uptake for the shortgrass steppe and 528,804 pixel-year time series for the northern mixed prairies. To produce estimates of 16-day C uptake for each day, we fit cubic spline regression models to the relationships between 16-day and cumulative C uptake (both absolute and relative % of total C uptake) and day of year for each pixel-year F I G U R E 2 Distributions of the two grassland ecoregions and the driest growing seasons between 2003 and 2020. Panel (a) shows the spatial distributions of the two ecoregions. Panel (b) shows the five most common driest years between 2003-2020 for each ecoregion. Panel (c) shows the spatial variation in the driest growing season, with lighter colors indicating more recent years. combination using the smooth.spline function in R. We then used the fit from each pixel-year spline model to produce smoothed daily estimates of C uptake for that pixel-by-year combination using the predict function in R. Note that each daily estimate for the cumulative C uptake time series represents the total amount of C uptake up to that day, while each daily estimate for the 16-day C uptake time series represents the total amount of C uptake in the preceding 16 days. To represent the 'non-drought' C uptake time series for a pixel, we used the median value for each day from among all 17 of years data that excluded the driest year. The remaining year was then the driest year on record for the pixel.
To determine how the time series of cumulative and 16-day C uptake differed between non-drought growing seasons and drought, we compared the drought time series to the 17-year median series.
We focused on daily absolute and percent (drought-median/median) differences between drought and non-drought years for both the 16-day and cumulative C uptake time series. For each day, we quantified the interquartile range of these estimated differences to estimate spatial variability in drought impacts across each ecoregion.
We repeated this analysis for NDVI.

| Drought impacts to vegetation phenology
To determine how the timing of C uptake was impacted by drought, we focused on the impacts to the day by which 25%, 50%, and 75% of total, cumulative growing season C uptake occurred. For example, if 50% of total C uptake is reached earlier during drought, that would indicate that C uptake becomes accelerated or more concentrated earlier in the growing season. By contrast, if 50% of total C uptake is achieved later during drought, that would indicate C uptake becomes delayed or more concentrated later in the growing season.
We estimated these shifts (in days) for each pixel by comparing the relativized time series of C uptake (% of total C uptake vs. day) between the drought year and the 17-year median non-drought time series, specifically by subtracting the Julian day for each drought year from the Julian day of the median non-drought time series for our three metrics. We then estimated the spatial synchronization, that is, autocorrelation, in these shifts. We first tested whether spatial autocorrelation was present using a two-sided Moran's I autocorrelation index using the ape package in R. If significant (p < .05) spatial autocorrelation was detected, we then performed variogram analysis using the gstat package in R to characterize that autocorrelation. We specifically focused on the range of autocorrelation in the variogram model, or the distance at which spatial autocorrelation saturated.
To assess whether the mean of a given distribution statistically differed from zero, we conducted bootstrapped (randomized with replacement, n = 100 per iteration, 1000 iterations) one-and two-sided t-tests (one sample), extracting the T statistic from each iteration. If the 99% confidence intervals of the bootstrapped T statistic distribution failed to overlap with zero, we deemed the mean estimate of the distribution statistically different from zero (Felton et al., 2021). We applied the same approach to test whether the distributions for a given metric statistically differed between the two ecoregions but instead used two-sample Kolmogorov-Smirnov tests, a nonparametric test of the equality of distributions. In this case, we extracted the D statistic (distance) from each iteration.

| Predicting peak reductions in vegetation productivity
We also considered seasonal changes in weather as potential drivers of peak reductions in vegetation productivity during drought. We first quantified and compared changes in the seasonality of temperature, precipitation, and VPD during drought to the 17-year nondrought average (mean) for all pixels. We split the growing season into meteorological spring (March-May) and summer (June-August) following Swain and Hayhoe (2015). For each pixel-year-season combination, we summed daily precipitation and calculated the average of daily temperatures and daily maximum VPD across each season.
For each pixel, we then assessed how drought altered both absolute and relative (%) changes (drought-mean/mean) in seasonal temperature, precipitation, and VPD for spring and summer. As another metric of changes in precipitation seasonality, we assessed how the percent of spring precipitation (as a share of spring + summer precipitation) changed during drought, relative to the 17-year average.
We then studied how the spatial variation in peak NDVI reduc-

| Drought overview
Between 2003-2020, the three most common driest years across the shortgrass steppe were 2011 (48.3% of pixels), 2012 (34.2% of pixels), and 2020 (12.6% of pixels), while across the northern mixed prairies, they were 2012 (41.5% of pixels), 2017 (25.2% of pixels), and 2020 (14.2% of pixels) (Figure 2b,c). During the driest years, growing season precipitation was reduced by a median of 57.4% (−215.4 mm) across the shortgrass steppe and 47.8% (−189.8 mm) across the northern mixed prairies (Table S1). Growing season temperature increased by a median of ~8% across both ecoregions during the driest years, representing a median increase of 1.4°C across the shortgrass steppe and a 1°C for the northern mixed prairies (Table S1).

| Timing, magnitude, and progression of drought impacts
Absolute rates of C uptake during drought were similar to median rates during spring for both ecoregions and were even enhanced by drought across the northern mixed prairies up to an estimated median peak increase of 8.7 g C m −2 16-days −1 during May (Figure 3; Table S2). C uptake became progressively reduced by drought during late spring in the shortgrass steppe, but not northern mixed prairies, with peak reductions occurring during early summer in both ecoregions. The same patterns were observed for the 1 km subset ( Figure S4) and for NDVI responses ( Figure S5). The median absolute peak decrease in C uptake across the shortgrass steppe was −22.1 g C m −2 16-days −1 , while across the northern mixes prairies it was −25.6 g C m −2 16-days −1 (Figure 3; Table S2). Although the median magnitude of absolute peak reductions in C uptake were comparable between ecoregions, they translated to greater relative (%) reductions across the shortgrass steppe (−65.6%) than the northern mixed prairies (−49%) ( Figure S6). Overall, absolute peak reductions in C uptake were greater than increases during drought. While absolute reductions in C uptake were comparable in magnitude among the two ecoregions, this translated to greater relative reductions across the shortgrass steppe ( Figure S7).
Median estimated peak decreases in 16-day C uptake occurred on 30 June in the shortgrass steppe and 6 July in the northern mixed prairies. Because estimates for each day reflected the total C uptake during the preceding 16 days, peak reductions in C uptake during drought primarily occurred during mid-and late June across both ecoregions. Probability density functions of each ecoregion showed considerable overlap in the relative likelihoods of these days, but the highest likelihoods indicated earlier peak decreases across the shortgrass steppe (Figure 4). There was low spatial variability in this timing, particularly across the northern portions of the northern mixed prairies. An exception was in the southern portions of the shortgrass steppe, where the effect of July-September monsoonal rains-which deliver roughly half of total annual precipitation-likely pushes the drought response window later Notaro et al., 2010). This resulted in a second peak in the probability density function on 18 August in the shortgrass steppe, meaning peak reductions in C uptake during drought occurred in early August across the southern-most portions of the shortgrass steppe ecoregion.
Cumulative C uptake during drought slowed early in the growing season as compared to median (non-drought) rates across the shortgrass steppe (Figure 5a). By contrast, cumulative C uptake across the northern mixed prairies did not differ considerably between drought and F I G U R E 3 Intra-annual dynamics of drought impacts to carbon uptake within two grassland ecoregions. Panels depict the estimated change in 16-day total carbon uptake throughout the growing season during drought relative to the median carbon uptake (dashed black line) in (a) the shortgrass steppe and (b) the northern mixed prairies. In both panels, the white line indicates the median estimated change in carbon uptake across each ecoregion for that day, while the black shading indicates the interquartile range.
non-drought years (Figure 5b). The same pattern was observed for the subset of pixels analyzed at 1 km resolution ( Figure S8). Thus, the median reduction in total C uptake at the end of the growing season was significantly larger across the shortgrass steppe (−142.8 g C m −2 ) than the northern mixed prairies (−71.9 g C m −2 ) (D = 0.41 ± 0.005; Table S2).
On a relative (%) basis, the reductions in C uptake were also larger across F I G U R E 4 Timing of absolute peak reductions in grassland carbon uptake during drought. The main panel shows the spatial variation in the day at which maximum reductions in carbon in uptake occurred (in g C m −2 16-days −1 ). The inset panel show the probability density functions (scaled to 1) to highlight the relative likelihoods of these days across the two ecoregions.

F I G U R E 5
Temporal progression of cumulative carbon uptake throughout the growing season in non-drought versus drought years in (a) the shortgrass steppe and (b) the northern mixed prairies. In both panels, the solid lines depict the median carbon uptake across sites in nondrought (black line) and drought years (red line), while the shading around each line depicts the interquartile range across sites (spatial variation). the shortgrass steppe and showed a stronger spatial pattern of increasing % reductions from north to south (Figure 6), suggesting increased losses in total C uptake during drought from north to south ( Figure S9).
Overall, reductions to cumulative C uptake during drought occurred earlier and for longer across the shortgrass steppe compared to the northern mixed prairies, resulting in both greater absolute and relative losses in total C uptake by the end of the growing season.

| Drought impacts to vegetation phenology
Drought advanced the day by which 50% of total growing season C uptake occurred across 90% of pixels (Figure 7). The advancement was significantly greater across the shortgrass steppe (median = −22 days) than the northern mixed prairies (median = −12 days) (bootstrapped D = 0.45 ± 0.005; Table S3).
There was also significant spatial autocorrelation in these shifts within both ecoregions (Moran's I p < .001) and variogram models estimated the range of autocorrelation to be 134 km across the shortgrass steppe and 485 km across the northern mixed prairies, implying regional-scale synchrony. Qualitatively similar results were observed earlier in the growing season for the day by which 25% of total C uptake occurred ( Figure S10; Table S3). By contrast, in the later growing season, the impacts to the day by which 75% of total C uptake occurred were weaker and less consistent (Table S3). Spatial variation in the day by which 75% of total C uptake occurred was negatively correlated with latitude (Spearman's ρ = −.24, p < .001, Figure S11). This relationship was opposite for the day by which 50% (ρ = .34, p < .001) and 25% (ρ = .49, p < .001) of total C uptake occurred.

| Predicting peak reductions in vegetation productivity
While drought was consistently associated with reduced precipitation and increased temperature and VPD during both spring and summer, the magnitude of these effects varied between seasons and ecoregions (Supporting Information). In general, the absolute magnitude of decreases in precipitation and increases in temperature and VPD were greatest during summer (Table S4; Figure 8 inset) and thus coincided with peak reductions in C uptake (Figures 3 and 4) and NDVI ( Figure S5). Reductions in summer precipitation and increases in VPD were each significant predictors of spatial variation in peak NDVI reductions during drought, but the model with VPD explained roughly 10% more variation (Figure 8; Table S6). The slopes of these relationships varied significantly by ecoregion and were greater in the northern mixed prairies, implying greater sensitivity of vegetation greenness to VPD increases across the cooler northern mixed prairies.

| DISCUSS ION
Our most important finding is that, over the first two decades of the 21st century, drought-induced reductions to western US grassland F I G U R E 6 Relative impacts of drought on total grassland carbon uptake. The main panel map depicts spatial variation in the percent change in total carbon uptake during the driest year relative to non-drought years. The inset depicts probability density functions (scaled to 1) that show probabilities of these changes in each ecoregion.
C uptake consistently peaked during early summer in mid-and late June (30 June in shortgrass steppe and 6 July in northern mixed prairies; Figure 3). This response was consistent spatially across the two expansive grassland ecoregions and temporally across different drought years. The only exception was in southern portions of the shortgrass steppe, where a monsoonal climate delays peak reduction of C uptake to August (18 August; Figure 4b). The timing of these impacts suggests that commonly-observed losses in annual grassland productivity due to drought in this region (Hoover et al., 2014;Knapp et al., 2015) are driven primarily by early summer reductions in C uptake (Figure 3). Spatial variation in the magnitude of peak reductions in vegetation greenness during drought was related to increases in temperature and VPD, which reflect an increase in atmospheric demand for water loss from plants ( Figure 8). The timing of reductions in C uptake during drought across the expansive western US Great Plains biome is likely constrained and related to the timing of both plant water supply and demand. In the future, combining our results describing spatiotemporal effects of drought on C uptake with high-resolution remote-sensing based measures of water-use efficiency could provide mechanistic insight into the underlying ecophysiological processes (Cooley et al., 2022).

| Intra-annual dynamics of drought impacts to carbon uptake
We originally speculated that reductions to grassland C uptake due to drought could increase, decrease, or be consistent throughout the growing season (Figure 1). The drought response pattern of C uptake across the western US Great Plains biome we observed was defined by limited effects early in the growing season and maximum reductions during the mid-growing season F I G U R E 7 Drought impacts to the day by which 50% of total grassland carbon uptake is reached. Both the map and inset density plot depict the difference (days) between the day by which 50% of total cumulative carbon uptake is reached between the driest year and the long-term median. Negative values indicate this day is reached earlier during drought.

F I G U R E 8
Relationship between the magnitude of peak reductions in Normalized Difference Vegetation Index (NDVI) and the magnitude of changes in summer vapor pressure deficit during years with the greatest precipitation deficits. Inset shows the average change in summer versus spring vapor pressure deficit for each ecoregion. Error bars denote the standard error of the mean and the R 2 is from the 'full' model across both ecoregions. VPD, vapor pressure deficit.
( Figure 3). The pattern most closely resembles an intensification of impacts through time (Figure 1 red line) but is more nuanced: a 'U-shaped' response with peak reductions occurring in June.
This likely reflects the phenological balance between low leaf area during the spring and leaf senesce in the fall, resulting in a relatively constrained timing of sensitivity to water stress . Our observations are broadly consistent with a previous continental-scale analysis of the 2012 US drought, which showed reductions to C uptake were primarily restricted to the summer months across much of the continent (Wolf et al., 2016).
By analyzing a larger collection of drought years, we show this drought response pattern is highly consistent across the western Great Plains.
Revealing the intra-annual response patterns of C uptake to drought across a large domain adds clarity to often-observed reductions in annual net primary productivity within grasslands. While spring stimulation to C uptake during drought, likely via enhanced temperatures, has the potential to offset summer C losses in some ecosystems (Wolf et al., 2016), such as in forests (Kljun et al., 2006;Wolf et al., 2014), spring C uptake was not stimulated across the shortgrass steppe. By contrast, C uptake was enhanced in the northern mixed prairies (Figure 3), likely through stimulated vegetation growth given the surprisingly large magnitude of spring warming in this ecoregion (Table S5). But this stimulation in C uptake was minimal in magnitude relative to the subsequent summer reductions.
Consequently, total C uptake in both these grassland ecoregions was reduced by drought. Early-to-mid-summer reductions to C uptake appear to drive annual-scale losses in total C uptake across these grasslands (Figures 3 and 5).
Vegetation phenology offers one explanation for the pattern of C uptake responses to drought. Cumulative C uptake is higher during summer than spring for both ecoregions in non-drought years: a median of 53% greater in shortgrass steppe and 129% greater in northern mixed prairies. Thus, grassland C uptake is most responsive to fluctuations in water availability when rates of C uptake are typically highest. While perhaps intuitive, an experiment in European grassland showed the opposite pattern, such that periods of high vegetation growth buffered drought effects on overall C uptake (Hahn et al., 2021). Moreover, a considerable fraction of vegetation productivity can occur during the spring months in these ecoregions (Smart et al., 2021) and we found that the fraction of total C uptake during spring was increased by drought. Median spring C uptake was 40% greater than summer during drought in the shortgrass steppe and median summer C uptake was only 38% higher than spring C uptake in the northern mixed prairies. This connects to our observation that the day by which 25% ( Figure S10) and 50% (Figure 7) of total C uptake occurs was consistently advanced by drought and suggests this advancement was primarily driven by reductions in C uptake during summer rather than by increases during spring (Figure 3). Thus, increased fractions of total C uptake occurred earlier in the growing season during drought because C uptake was reduced primarily during the summer months.
If C uptake is most sensitive to drought during the early-to midsummer, then perhaps it is also most sensitive to moisture pulses during those same periods. Both the relatively shallow rooting depth of perennial grasses (Fan et al., 2017;Nippert et al., 2012) and the high proportion of summer precipitation (Lauenroth et al., 2014) across the Great Plains suggests vegetation in these ecosystems should be sensitive to summer moisture pulses. In fact, leaf physiology and resultant C uptake in the western US Great Plains can be sensitive to precipitation events at daily timescales (Petrie et al., 2016;Sala & Lauenroth, 1982). This highlights the importance of understanding how fine-scale fluctuations (e.g., subweekly) in water availability impact grassland functioning. Water cycle intensification includes not only more frequent droughts, but also increases in precipitation event size and intensity (Ficklin et al., 2022). Larger precipitation events in grasslands could 'rescue' ecosystem functioning from drought, but likely only if their timing is right, as field experiments in the shortgrass steppe suggest (Hoover et al., 2022;. Future efforts to understand the implications of increasing climatic variability should thus also aim to focus on when ecosystem processes are most sensitive to both positive and negative fluctuations in water availability within a growing season.

| Seasonality of water stress
The seasonality of C uptake responses may also be due to the seasonality of water limitation. Counterintuitively, droughts in the Great Plains increase the abundance of cool-season grasses , implying that water limitation may peak in the summer months when warm season (C 4 ) grasses-presumably well-equipped for arid conditions (Sage, 2004)-are most active. This compositional change indirectly supports our observation of peak C uptake losses during the summer months when C uptake of warm season grasses should be greatest (Figure 3). One mechanism for this seasonal response pattern could be a relative shift toward more spring precipitation during drought. However, we did not find a large or consistent change in the percent of spring precipitation ( Figure S11). This does not exclude other indices of precipitation seasonality, such as the slope of the relationship between monthly temperature and precipitation Lauenroth et al., 2014), from suggesting a shift in precipitation seasonality during drought. In fact, we suspect that because the highest temperatures and largest absolute reductions in precipitation occurred during the summer months, a decrease in the monthly precipitation-temperature slope, and thus a reduction in 'summer precipitation dominance,' likely occurred during droughts across this biome.
However, we suggest a more general explanation for the intraannual pattern of drought response is the seasonality of water stress arising from both water supply and demand. Of course, precipitation supply was reduced in both seasons and tended to be reduced to a greater absolute magnitude during the summer (Table S4). However, there were also clear differences in how drought impacted absolute water demand between the two seasons; the median drought-induced increase in maximum VPD was 2-3 times greater during the summer than spring (Table S4). The relationship between VPD and peak reductions in NDVI (Figure 8) points to the role of rising VPD in reducing stomatal conductance, photosynthesis, and transpiration (Fu et al., 2022;Grossiord et al., 2020;Yuan et al., 2019).

| Limitations
While our analysis demonstrates patterns of C uptake over time and space with high fidelity given the effects of drought inferred by precipitation amount, it does not directly compare C uptake with plant available water (e.g., soil moisture) or the drivers of plant water demand (e.g., the inextricable effects of temperature and vapor pressure effects). This is important because the majority of root biomass in Great Plains grasses is in shallow soils layers (< 50 cm), which are subject to high temporal variability in soil moisture during the growing season (Lauenroth et al., 2014). And increasing soil moisture variability alone, without associated changes in total precipitation or mean soil moisture, can significantly alter net primary production across the Great Plains (Knapp et al., 2002;Thomey et al., 2011;Wilcox et al., 2015). With the continued maturation of remotelysensed soil moisture products (e.g., those based on microwave emissions; Feldman et al., 2022), it will become possible to incorporate high-resolution, spatiotemporal patterns of soil moisture dynamics into studies of productivity.
Plant water demand, which can be inferred through VPD, has also been demonstrated to have a significant impact on grassland function. Given that VPD is calculated as a function of both temperature and vapor pressure, we cannot distinguish comparatively between their effects. However, increasing temperature drives both a higher driving gradient for water loss through transpiration, as well as introduces the possibility of effects on respiration and direct damage to photosynthetic machinery that reduces primary productivity (Kirschbaum, 2004). Nevertheless, it is arguably the combination of hot and dry conditions that will produce the largest reductions in ecosystem functioning during drought.
Finally, our analysis also does not consider other concurrent changes occurring during the study period, including rising atmospheric CO 2 , nitrogen deposition, or vegetation compositional changes. These factors will interact with future drought events in complex ways. For example, rising atmospheric CO 2 could ameliorate drought impacts on C uptake by reducing plant water demand (Swann et al., 2016). Simultaneously, ongoing and widespread greening across the Northern Great Plans, often in the most arid locations (Brookshire et al., 2020), could exacerbate drought impacts on C uptake through increased plant water use associated with higher leaf area (Chen et al., 2022). The pace of these changes and the com-

| Implications
Both ecoregions showed an intensification of drought-induced reductions to C uptake through time that peaked in June (Figures 3 and   4). Time-sensitive management decisions, such as when and where to graze cattle, may be aided by the knowledge that reductions in vegetation productivity during the summer are all but guaranteed during drought (Figure 3). By contrast, the spring and fall months may exhibit more typical levels of productivity, indicating that the shoulder seasons of spring and fall can serve as drought refuges for ecosystem services in grasslands, such as forage production. Whereas C uptake may be limited by low leaf area or senescence during these shoulder seasons, projections also indicate increased grassland productivity during these seasons due to increases in growing season length with climate change (Hufkens et al., 2016). Management may also be aided by the knowledge that as drought conditions emerge, one can be expected to reach 50% of total production 2-3 weeks earlier during dry years (Figure 7), potentially advancing timelines of decision calendars developed for more typical years. However, these shifts varied by ecoregion and displayed regional-scale spatial synchrony (autocorrelation), implying a 'one-size fits all' approach may not be appropriate for management. Still, the large scale of spatial synchrony in these shifts-134 km in the shortgrass steppe and 485 km in the northern mixed prairies-suggest extensive regional boundaries may be a reasonable scale of decision making across this biome when factoring in the timing of drought impacts.
Our results serve as an example of the potential spatial variation in climate vulnerability across the western US Great Plains biome.
C uptake across the shortgrass steppe was reduced earlier, longer, and to an overall greater magnitude than the northern mixed prairies (Figures 3, 5 and 6). Recent work further suggests that even if the structure and functioning of the shortgrass steppe kept pace with climate change throughout the 21st century, many areas would still experience losses in net primary productivity. The same scenario projected widespread increases in productivity across the northern mixed prairies (Felton et al., 2022). Our results indicate greater vulnerability of the southern than northern plains to drought intensification across the western US (Bradford et al., 2020).

ACK N OWLED G M ENTS
This project was supported by a USDA-NIFA Fellowship (Award # 2021-67034-35121) awarded to A. Felton. Discussions with ranchers within the Great Plains focused our thinking about the connection of these results to management implications. We thank two reviewers for constructive comments on the manuscript.

CO N FLI C T O F I NTE R E S T S TATE M E NT
The authors declare no conflicts of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The core spatiotemporal datasets and the associated R code used for analyses and figure production are publicly available through a Zenodo repository (https://doi.org/10.5281/zenodo.6977486).