Introduction

A key climate change prediction is an increase in the interannual variability of precipitation, with consequent increases in the frequency and severity (or extremity) of dry and wet years (IPCC 2012). Such changes to precipitation regimes are expected to differentially impact terrestrial carbon cycling dynamics (Ciais et al. 2005; Knapp et al. 2008; Reichstein et al. 2013; Shi et al. 2013; Poulter et al. 2014; Ahlstrom et al. 2015; Frank et al. 2015). Indeed, there is growing evidence that ecosystem processes, such as net primary production (NPP), may be more sensitive to dry versus wet years or vice versa (Knapp and Smith 2001; Wu et al. 2011; Haverd et al. 2016; Wilcox et al. 2017). However, current understanding of how precipitation affects ecosystem processes does not fully incorporate the potential for such asymmetric responses, particularly in response to extremes. As a consequence, it is evident that our ability to project how ecosystem processes will respond to predicted changes in precipitation is constrained by a limited understanding of the impacts of climatically extreme dry and wet years.

Current understanding of relationships between precipitation and NPP, specifically the aboveground fraction of NPP (ANPP), is derived from empirical models based on either spatial or temporal data. Spatial models relate mean ANPP across sites to natural gradients in mean precipitation (e.g., from deserts to forests; Huxman et al. 2004), while temporal models relate within-site variation in ANPP to interannual variation in precipitation over multiple years (Lauenroth and Sala 1992; Sala et al. 2012). Spatial models not only yield steeper slopes than temporal models (Sala et al. 2012), but also often yield a nonlinear, saturating response of ANPP to precipitation when incorporating broad precipitation gradients and multiple biomes (Huxman et al. 2004). By contrast, temporal models consistently yield a linear relationship between ANPP and precipitation within sites (Huxman et al. 2004; Sala et al. 2012; Estiarte et al. 2016; Knapp et al. 2017a), with nonlinearities often weak (Hsu et al. 2012). Spatial models reflect long-term climatic influences on ecosystem properties, and thus necessarily incorporate large gradients in soil texture, soil nutrients, and vegetation structure into their predictions. As a result, such models may not be appropriate for projecting responses of any one ecosystem to changes in precipitation, for instance, due to site-specific vegetation (e.g., tiller density; Yahdjian and Sala 2006) or biogeochemical constraints (e.g., soil nitrogen; Huxman et al. 2004). Current temporal models have also been recently demonstrated as preferable over spatial models for benchmarking ecosystem models (Estiarte et al. 2016). Yet, despite a current preference for temporal models, evidence has accumulated that ANPP responds asymmetrically to wet versus dry years, which has been suggested to potentially conflict with the linear predictions of these models (Knapp et al. 2017a).

The concept of asymmetry signifies that primary production may harbor differential sensitivities to increases versus decreases in precipitation, relative to production during years of nominal precipitation amounts (Knapp et al. 2017a). Such asymmetry may be either positive or negative. Positive asymmetry denotes that the relative increase in ANPP is greater during wet years than declines during dry years, and may arise if drought-buffering mechanisms (e.g., physiological adjustments) operate in concert with high sensitivity of plant growth to wet years (Knapp and Smith 2001; Wilcox et al. 2017), or if the cost of deploying new tissues during resources pulses is less than abscission during stress (Sala et al. 2012). By contrast, negative asymmetry denotes a greater relative sensitivity to drought, and may occur if resource co-limitation (Huxman et al. 2004) or hypoxia limits production in wet years and water stress during drought leads to continual declines in rates of ecosystem processes (Knapp et al. 2017a). Although current understanding suggests a greater prevalence of positive asymmetry (Knapp and Smith 2001; Wu et al. 2011; Haverd et al. 2016; Wilcox et al. 2017) relative to negative asymmetry (Zscheischler et al. 2014; Zhang et al. 2017), much of our understanding is derived from responses of aboveground NPP to precipitation (Knapp and Smith 2001, but see Wilcox et al. 2017) in water-limited ecosystems (e.g., Haverd et al. 2016). This calls for both comparisons and an integration of above- and belowground carbon exchange responses to precipitation variability (Wilcox et al. 2017), as differential sensitivity of carbon inputs versus losses to precipitation change can have consequences for ecosystem carbon balances.

One key consequence of asymmetric responses of ANPP to precipitation variability is potential divergence from the consistently linear prediction of current temporal models (Huxman et al. 2004; Sala et al. 2012), such as towards a nonlinear, saturating relationship (Knapp et al. 2017a; Luo et al. 2017; Wilcox et al. 2017; Zhang et al. 2017). This divergence may be particularly pronounced when both dry and wet precipitation deviations become extreme (Knapp et al. 2017a), such as at or beyond the 1st and 99th percentiles of the historical record (Smith 2011; Kayler et al. 2015). Given that temporal records of precipitation and ANPP are often limited in duration and thus their representation of climatic variability, such models often do not include ecosystem responses to extreme values in precipitation. As a result, current understanding of ANPP–precipitation relationships is largely derived from responses to nominal precipitation variability. Because of such limitations, novel approaches to assess the sensitivity of ecosystem processes to future changes in precipitation that include both dry and wet extremes must be considered (Smith 2011; Knapp et al. 2017b).

To address these limitations, we employed a novel experimental approach within a mesic ecosystem (Fig. 1). Over a single growing season, we exposed an intact, mesic grassland to an experimental gradient of precipitation that specifically incorporated extreme values of both precipitation increases and decreases. What constitutes an extreme climate event has been defined in several ways (Van de Pol et al. 2017), specifically through climatological (IPCC 2012), organismal (Gutschick and BassiriRad 2003), and ecological (Smith 2011; Hoover et al. 2014) perspectives. For our purposes, we employ a climatological definition of climatic extremity, specifically through a statistical perspective that focuses on percentiles of the long-term climate record, such that the 5th precipitation percentile (drier than 95% of years on record) and below constitutes extreme drought, while the 95th precipitation percentile (wetter than 95% of years on record) and above constitutes an extreme wet year. Previous precipitation manipulation experiments imposing such types of extremes have often employed a limited number of treatment levels (e.g., drought and control; Hoover et al. 2014), but here we utilized a replicated regression experimental design (Cottingham et al. 2005; Gotelli and Ellison 2013) to produce a precipitation gradient that ranged from the 1st and 99th growing season precipitation percentiles of the 110-year record. This precipitation gradient consisted of 11 levels with four replicates per each level, akin to an 11-year temporal dataset (Fig. 1b). As a consequence, our experimental design allowed us to test—for the first time—the sensitivity of an intact ecosystem to an extensive gradient of 11 levels of precipitation within a single growing season that specifically included extreme values in precipitation.

Fig. 1
figure 1

a The 2016 field experiment at the Konza Prairie Biological Station in northeastern Kansas. b Relationship between growing season (May–August) precipitation percentiles from the long-term record and total growing season precipitation amounts added. Specific emphasis was placed on the distributional tails of the long-term record. Inset is the relationship between precipitation amounts and mean soil moisture availability during the growing season, which resulted in linear relationships (R2 = 0.66). Standard errors on each point are plotted yet too small to be visible

Full size image

Our goals were to (1) assess whether the magnitude of impacts on carbon inputs (ANPP) versus losses (soil respiration; Rs) differed between extreme negative versus positive deviations in precipitation, relative to nominal precipitation amounts (evidence for asymmetry), and (2) test the temporal model prediction that ANPP, and additionally Rs, exhibits a linear relationship with precipitation, with an emphasis on the distributional tails, i.e., extremes, of the 110-year precipitation record. We chose to assess Rs in addition to ANPP because this process constitutes the largest flux of carbon from terrestrial ecosystems to the atmosphere (Liu et al. 2016), and because such relationships of Rs with precipitation and soil moisture remain largely uncertain in current ecosystem models (Hawkes et al. 2017). This further allowed us to gain inference about how sensitivity to precipitation change may differ between above versus belowground carbon exchange processes (Liu et al. 2016; Wilcox et al. 2017), and processes governing ecosystem carbon inputs (ANPP) versus losses (Rs).

Because previous findings of positive asymmetry in ANPP are derived primarily from responses to nominal precipitation variability (Knapp and Smith 2001, but see Wilcox et al. 2017), we expected to see evidence of negative asymmetry of ANPP emerge after incorporating both wet and dry extremes into our precipitation gradient, as predicted by Knapp et al. (2017b). We expected these responses to be driven by the dominant control of soil moisture availability under extreme drought (Hoover et al. 2014), and emergent constraints of resource co-limitations, such as light, under extreme wet conditions (Huxman et al. 2004). In concert with negative asymmetry, we hypothesized a nonlinear, saturating relationship would best explain the response of ANPP to precipitation. By contrast, we expected to see near symmetrical sensitivity of Rs to extremes driven by consistently negative effects of drought and waterlogged soils. Specifically, due to negative effects related to increasingly reduced soil pore connectivity and substrate movement under drought, and reduced oxygen diffusion under saturated soils, we expected to see evidence of an underlying hump-shaped response of Rs to precipitation emerge, consistent with current theory (Moyano et al. 2013).

Materials and methods

Experimental site

This experiment was conducted at the Konza Prairie Biological Station during the 2016 growing season (May–August, Fig. 1a). The Konza Prairie is a 3487-ha biological reserve of native unplowed tallgrass prairie located in the Flint Hills region of northeastern Kansas. This region is characterized by warm, wet growing seasons and cold, dry winters. Mean annual temperature at the Konza Prairie is ~ 13 °C, while mean annual precipitation is ~ 892 mm, of which that vast majority occurs during the growing season. We selected an upland, annually burned watershed with shallow fine-textured Florence soils (bulk density of ~ 1 g cm−3, see Ransom et al. 1998 for specific soil details) to conduct the experiment. Mean daily potential evapotranspiration during the manipulation period was 4.89 mm day−1 (AWE01 Konza LTER dataset). The plant community composition of this upland site is dominated by rhizomatous perennial warm-season C4 grasses, in particular the species Andropogon gerardii, Sorghastrum nutans, and Schizachyrium scoparium, with lesser abundant forb and graminoid species contributing to plant species richness (Knapp et al. 1998).

Experimental design and treatments

We used a replicated regression experimental approach (Cottingham et al. 2005; Gotelli and Ellison 2013) within a randomized block design (n = 2 blocks, n = 2 replicates per treatment per block) to produce a gradient of growing season precipitation amount. In these experimental designs, statistical power is generated primarily from the treatment gradient, and thus the number of treatment levels, rather than the number of replicates per each treatment level (Cottingham et al. 2005). However, adding replication to each precipitation level (n = 4) provided increased statistical power to our experimental design and subsequent analyses. Each precipitation level corresponded to a specific percentile of the 1900–2012 daily precipitation records for Manhattan, KS from the NOAA National Climatic Data Center (now Global Historical Climatology Network) and additionally the Kansas State University Weather Data Library (see Hoover et al. 2014). Precipitation levels ranged from the 1st (driest) to the 99th (wettest) percentiles of the 112-year record. Percentiles were calculated for total precipitation amounts from the time period May 7th–August 31st using the quantile function in base R. The precipitation gradient ranged from under 200 mm to over 1000 mm, with 11 levels in total (Fig. 1b).

The target precipitation percentile amounts were achieved utilizing 100% rainfall exclusion shelters deployed on a large scale (Fig. 1a). Polycarbonate roofs (2.2 × 2.2 m) were mounted onto four fence T-posts to exclude all ambient rainfall. Shelter roofs were sloped from a height of 100 cm to ~ 70 cm to allow for rainfall drainage into gutters attached to tubing, which shunted rainfall away from the study site. Experimental plots were 2 × 2 m with a nested 1 × 1 m sampling plot, which produced a ~ 0.7 m buffer zone. Experimental plots were spaced 2 m apart, with all shelters sloped in the same direction (west) against the prevailing winds. We simulated rainfall events on a weekly basis by conducting water additions, with amounts varying each month according to monthly differences in precipitation amount based on the long-term 112-year record. For each month, rainfall events were divided into four equal rainfall applications that ocurred on a weekly basis. Monthly amounts were calculated by multiplying the long-term monthly mean by the percent difference of each growing season precipitation percentile from the growing season mean.

Water additions were performed in the evening (after 17:00) to ensure effective treatment application through minimizing evaporative losses and avoiding confounding factors associated with mid-day (e.g., 11:00–14:00) conditions such as high solar radiation (up to 2000 photosynthetic active radiation) and peak daily temperatures. Water additions were accomplished utilizing a gas-powered water pump connected to nearby water tanks filled with water supplied from a local aquifer. We utilized flow meters (Great Plant Industries, Wichita, KS) to measure the amount of water to add to each plot, converting mm to gallons per 4 m2. All plots under the 50th precipitation percentile were watered over 1 day, while all plots in the 50th percentile and above were watered over 2 days to ensure effective treatment application by avoiding excessive surface runoff associated with the shallow soils of our upland site, which can range from 5 cm to 2 m in depth (Nippert et al. 2012). Nevertheless, experimental rainfall event sizes were comparable to that historically experienced in this system (Felton et al. in prep).

Soil volumetric water content was repeatedly measured during June–August to 20 cm depths, where the bulk of root biomass is located in upland sites within this ecosystem (Nippert et al. 2012). Soil moisture measurements were taken utilizing a portable Hydrosense II moisture sensor (Campbell Scientific, Logan UT), and were measured randomly in two different locations within each 1-m2 sampling plot once per week prior to water additions. The two measurements per sampling period were than averaged to yield a single soil moisture value per plot. Calibrations were performed by comparing probe values to known gravimetric water content values from soil samples within the site across a range of moisture levels following the methods of Dietrich and Smith (2016) and Wilcox et al. (2014).

We additionally assessed changes to light availability by measuring photosynthetic active radiation (PAR) above the canopy and at the soil surface throughout the 2016 growing season at three specific time points at the beginning of each month: June 6th, July 7th, and August 3rd. PAR was measured over 1 day when conditions were clear between 11:00 and 13:00 utilizing a portable ceptometer (Decagon Devices, Pullman, WA, USA). We measured at three locations at the soil surface within each 1-m2 sampling plot and once above the canopy, yet below the shelter. We then divided the average of the below-canopy values by the above-canopy value to provide a single value of % below-canopy PAR infiltration.

Carbon exchange responses

Aboveground net primary production was estimated at the end of the growing season in early September of 2016. Due to this site being burned annually every spring, assessing ANPP provided strong inference about the impacts of the treatments on primary production. To estimate ANPP, all aboveground biomass was clipped at ground level within two 0.1-m2 quadrats located in the center of each 1-m2 sampling plot. The two measurements for each plot were then averaged. Aboveground biomass was field-sorted to graminoid, forb, and woody plant functional groups. Biomass was then immediately dried at 60 °C for 48 h and weighed to the nearest 0.01 g. Woody biomass, which represented a small proportion of total ANPP (on average ~ 3%), was excluded from data analysis.

Measurements of in situ Rs were taken every 2 weeks in June, July, and August of 2016. A LI-8100 portable gas exchange system (LI-COR Biosciences, Lincoln, Nebraska) was used to take measurements from two locations in each plot. At the beginning of the growing season (late April), two PVC collars (10 cm diam.) were installed diagonally in opposite corners within each nested 1-m2 sampling plot. Measurements occurred at least 3 days after water additions. On the day prior to measurements, any aboveground plant biomass inside the collars was clipped and removed to isolate CO2 efflux. Flux values were recorded over a 45-s period once on each of the two collars within each plot, with measurements taken between 11:00 and 14:00. The two flux values were then averaged to yield a single flux value per plot for each sampling period. Soil moisture and temperature measurements were taken at the same time as Rs. Soil temperature was measured to 10 cm depths at the center of each sampling plot utilizing a thermocouple attached to the LI-8100 gas exchange system. Net photosynthetic responses of the dominant C4 grass Andropogon gerardii were additionally assessed throughout the growing season. Specific methods for net photosynthesis are described in the supplementary materials (ESM1 for detailed methods).

Plant community composition

Plant community composition was assessed in 2016 via visual percent cover estimates during early September. Percent cover was estimated to the nearest 1 or 5% for each species separately, following the methods of Koerner and Collins (2014).

Statistical analyses

We used two approaches to quantify asymmetry in carbon exchange responses to positive versus negative precipitation extremes. In line with Haverd et al. (2016), we defined asymmetry as

$$a_{\text{p}} = \left( {\bar{F}_{{ + {\text{p}}}} - \bar{F}} \right) - \left( {\bar{F} - \bar{F}_{{ - {\text{p}}}} } \right)/\bar{F},$$
(1)

where \(\bar{F}\) denotes the mean value of responses between the 15th and 85th precipitation percentiles (i.e., nominal variability), while \(\bar{F}_{{ + {\text{p}}}}\) denotes the average of the 95th and 99th percentiles and \(\bar{F}_{{ - {\text{p}}}}\) denotes the average of the 5th and 1st percentiles. A positive value indicates that the relative response was greater for extreme wet conditions (positive asymmetry), while a negative value indicates the relative response was greater for extreme dry conditions (negative asymmetry). To compliment this approach, we calculated Cohen’s D effect sizes (pooled standard deviation) for the same precipitation levels, quantifying the effect sizes for both wet and dry precipitation extremes.

To evaluate if carbon exchange responses to precipitation and soil moisture exhibited predominantly linear or nonlinear relationships, we tested whether linear or nonlinear least squares regression models best fit the responses of ANPP and Rs to both precipitation amount and soil moisture availability. The structure of the linear model used is the same as the least squares linear model (lm function in base R), whereas the basic structure of the nonlinear least squares model used (nls2 function in R, nls2 package) is

$$\gamma = \frac{{\theta_{1 } }}{{1 + \exp \left[ { - \left( {\theta_{2 } + \theta_{3} x} \right)} \right]}} + \varepsilon ,$$
(2)

where γ is the response variable (ANPP or Rs), x is the predictor variable, θ1 is a conservative approximation of the asymptote of γ (larger than any value in the dataset), and ε is the error term. θ2 and θ3 are starting values derived from an ordinary least squares regression of the logit transformation of the relationship between γ and x. We then used Akaike’s Information Criteria (AIC) model weights to provide inference on linear versus nonlinear model fits. We utilized a linear mixed effect model framework (lme function, nlme package) for model selection, which allowed us to treat plot nested within block as the random effect for both ANPP and Rs. We conducted backwards model selection (dropterm function, MASS package) and used AIC values as our selection criteria. For ANPP, the full model included mean soil moisture availability to 20 cm depths, light availability in August, and mean net photosynthesis of the dominant C4 grass A. gerardii. For Rs, the full model included soil moisture to 20 cm depths, soil temperature to 10 cm depths, and net photosynthesis of A. gerardii. Residuals were assessed for homoscedasticity of variance, while outliers and/or influential observation were assessed via outlier tests with Bonferroni-adjusted P values at the < 0.05 cutoff.

We used permutational multivariate analysis of variance with Bray–Curtis distance analyses and similarity percentages to assess plant community composition among previously defined categories of the precipitation levels. More specifically, these subsets included extreme dry (1st–5th percentile), nominal (15th–85th percentiles), and extreme wet (95th–99th percentiles). All analyses were conducted within the R statistical software environment (R Core Team 2018: R 3.2.1 GUI 1.66 Snow Leopard build).

Results

Mean soil moisture availability in the upper 20 cm of the soil profile increased linearly with precipitation amount (R2 = 0.66, Fig. 1b, inset). Light availability decreased considerably throughout the growing season and especially in the wettest plots (Figure S1), with < 5% of PAR reaching the soil surface in the 99th percentile plots by the beginning of August. Across the water availability gradient, ANPP values ranged from a minimum of 251.5 to a maximum of 760.9 g m−2, while Rs ranged from a minimum of 2.1 to a maximum of 13.7 μmol CO2 m−2 s−1. In terms of the extremes of the precipitation gradient, wet extremes produced a large positive effect on ANPP, while dry extremes resulted in a small negative effect (Fig. 2a). As a consequence, there was strong evidence in favor of greater relative ANPP responses to wet versus dry extremes and thus positive asymmetry to precipitation extremes (ANPP ap = 0.41, Fig. 2b). By contrast, Rs exhibited near equally large magnitudes of decreases and increases to dry versus wet extremes, respectively (Fig. 2a), with this process only slightly more responsive to dry versus wet precipitation extremes (Rsap = − 0.027, Fig. 2b).

Fig. 2
figure 2

The two approaches used to quantify asymmetric responses of ANPP and soil respiration to precipitation extremes. For both graphs, the mid-line represents values at nominal precipitation amounts (between 15th and 85th precipitation percentiles). a The magnitude of effect (Cohen’s D effect size) that each extreme had relative to values at nominal levels of precipitation variability. Error bars denote 95% confidence intervals, while gray lines denote the cutoff for ‘large’ effect sizes at ± 0.8. b The relative responses to extremes, such that a positive or negative values indicates responses were relatively larger for extreme wet (positive asymmetry) or dry treatments (negative asymmetry)

Full size image

A linear model was selected as best explaining ANPP responses to both precipitation (slope 0.40 g m−2 mm−1, intercept 251.78 g m−2) and soil moisture (slope 16.1 g m−2%VWC−1, intercept 31.02 g m−2), with both drivers explaining equal amounts of variation in this process (Fig. 3, Table S1). For Rs, model weights demonstrated consistent evidence in favor of a nonlinear model, with a saturating response to both precipitation (predicted asymptote 10.36 μmol CO2 m−2, inflection point 204.31 mm) and soil moisture (predicted asymptote 9.96 μmol CO2 m−2, inflection point 19.9%VWC, Fig. 4; Table S2). Stepwise selection resulted in soil moisture alone as the most parsimonious predictor of ANPP responses (Fig. 3b). For Rs, the interaction of soil moisture and temperature was selected as the best model, with a positive relationship of Rs to soil moisture (Fig. 4b), yet a negative relationship of Rs to increasing soil temperatures (Fig. 4d). This result is consistent with a highly significant interactive effect of temperature and soil moisture on Rs (P < 0.0001), and is reflected in the negative effect of soil moisture on soil temperature (R2 = 0.24, P < 0.001, Fig. 4c). Plant community composition did not differ among treatment subsets (Bray–Curtis dissimilarity, P = 0.70), and was driven by the abundances of the three C4 grass species that dominate throughout the site.

Fig. 3
figure 3

Response of aboveground net primary production (ANPP) to precipitation amount (a) and soil moisture availability (b) during the experiment. ANPP exhibited consistent evidence for a linear response to water availability, in terms of both precipitation and soil moisture, with a linear model supported over a nonlinear model in both cases (Table S1). R2 values were generated from linear models, and red trend lines depict the fit from a linear model. This figure appears in color in the online version of the journal

Full size image
Fig. 4
figure 4

Relationships among  soil respiration (Sr), precipitation, soil moisture, and soil temperature. a Relationship of mean values of soil respiration (June–August) with total growing season (May–August) precipitation amounts. Error bars denote standard the error of the mean. b Relationship of flux values with soil moisture availability to 20 cm depths. For both of these approaches, a clear nonlinear, saturating relationship emerged between soil respiration and water availability with a nonlinear model supported over a linear model (Table S2). Red trend lines depict the fit from a nonlinear model. c Significantly negative relationship between soil moisture and soil temperature (R2= 0.24, P < 0.001), which may be, in part, attributable to lower plant cover in drought plots (Fig. S1). d As a result of the negative relationships between soil temperature and soil moisture availability, a significantly negative relationship between soil temperature and Sr emerged (R2= 0.17, P < 0.001). This figure appears in color in the online version of the journal

Full size image

Discussion

Amid predictions and observations of changing frequencies and intensities of climate extremes (IPCC 2012), concerns have emerged that traditional approaches used to quantify relationships between climatic drivers and ecological processes may not be predictive of responses to climatic extremes with little or no historic analog (Kayler et al. 2015; Knapp et al. 2017a). Current models relating responses of ecosystem processes, such as ANPP, to variation in precipitation rarely include responses to precipitation extremes, and thus a pressing need has emerged for novel approaches to better understand the influence of extremes on previously defined relationships between precipitation and ecosystem processes. To address this, we experimentally imposed an extensive gradient of growing season precipitation amount and extremity within an intact mesic grassland ecosystem, and assessed responses of processes that govern both ecosystem carbon inputs (aboveground net primary production; ANPP) and losses (soil respiration; Rs). We sought to determine (1) the relative sensitivities of ANPP versus Rs to wet versus dry precipitation extremes, and (2) the consequences that responses to extremes have for the underlying relationships of these processes with precipitation. We report contrasting responses of ANPP and Rs to precipitation extremes, with strong evidence for positive asymmetry in ANPP responses to precipitation extremes, yet near symmetrical sensitivity in the response of Rs. Furthermore, while the responses of ANPP to both precipitation and soil moisture were consistently linear, Rs exhibited consistently nonlinear, saturating responses.

ANPP versus R s responses to precipitation extremes

Our hypothesis of near symmetry in the response of Rs to the precipitation gradient was supported, yet the pathway towards this symmetry differed from what we expected. We expected a unimodal relationship to underlie this symmetry, and thus for both wet and dry extremes to result in comparably negative impacts on Rs. Yet by contrast, a clear saturating relationship was observed (Fig. 4), with similar magnitudes of increases and decreases in Rs to wet and dry extremes, respectively (Fig. 2). Existing theory predicts that both wet and dry extremes should similarly negatively impact Rs, driven by different stresses that occur under extreme wet and dry conditions relative to an optimal balance of substrate and oxygen transport achieved at intermediate levels of water availability (Moyano et al. 2013). Specifically, whereas low soil moistures constrain substrate–microbe interactions to reduce heterotrophic and root respiration, saturated soil conditions may facilitate substrate mobilization to a degree that exceeds the physiological capacity of soil microbes. At the same time, saturated soils reduce aeration and produce frequent anoxic conditions that limit oxygen transport necessary for root and microbial respiration, further restricting Rs at the highest precipitation inputs. While we indeed find support for negative effects of drought on Rs (Fig. 2a), both the positive effects of wet extremes (Fig. 2a) and underlying saturation of Rs (Fig. 4) suggest that (1) the maximum response capacity of Rs to soil moisture was observed within our experiment, and (2) Rs may be relatively resistant to the negative effects of high soil moisture levels in this grassland.

Although our hypothesis of an emergence of negative asymmetry in ANPP responses to extremes (Knapp et al. 2017a) was not supported (Fig. 2), positive asymmetry is consistent with previous observations within this and many other grasslands (Knapp and Smith 2001; Wilcox et al. 2017). We suspect that the apparent greater sensitivity of ANPP to wet extremes may be driven by multiple, nonexclusive mechanisms. First, drought-buffering mechanisms that enhance precipitation use efficiency, such as previous-year soil moisture carryover, plant physiological adjustments, and their interaction may prevent strong declines in ANPP during a single dry year in mesic ecosystems, which appears to be supported here when we related ANPP to precipitation percentiles (Fig. S1b). At the same time, high growth rates of herbaceous vegetation, such as the C4 grasses dominant within our study region, matched with low light limitation through annual burning and high growing season temperatures (Taylor et al. 2017), all enable large aboveground growth responses to soil moisture surpluses. Such asymmetry may indeed differ within C3-dominated grasslands, particularly if the lower water-use efficiencies of this functional group (relative to C4 species) induces higher drought sensitivities. However, previous research suggests that such biotic controls of ecosystem responses to precipitation will be mediated by background climatic conditions, such as temperature (Wilcox et al. 2014; Taylor et al. 2017).

The differential sensitivity of above- versus belowground processes to precipitation change that we observed in our study is generally consistent with previous findings (Wu et al. 2011; Wilcox et al. 2014, 2017). For example, in a recent meta-analysis of grassland precipitation manipulation experiments, Wilcox et al. (2017) similarly found strong positive asymmetry in ANPP, yet symmetrical responses in belowground NPP (BNPP) to precipitation extremes. Greater sensitivity in ANPP versus BNPP to increases in soil resources has been attributed to shifts towards light limitation predicted by optimal allocation theory, of which there is empirical support (McConnaughay and Coleman 1999). However, while Rs and BNPP are both coupled with ANPP, Rs also necessarily integrates heterotrophic microbial responses to soil moisture changes, in contrast to BNPP, with Rs also likely responding to soil moisture fluctuations on finer timescales than BNPP. As a consequence, optimal allocation theory explaining ANPP versus BNPP responses may not fully apply to Rs, and thus our Rs results may rather reflect rapid metabolic responses to soil substrate and oxygen availability (Moyano et al. 2013). Nonetheless, the near symmetrical versus positive asymmetrical responses of carbon losses (Rs) versus inputs (ANPP) to precipitation extremes, respectively, suggests that the carbon balance of this grassland has the potential to shift towards a considerable sink during wet years.

Soil moisture as the central mechanism

We found consistent evidence for changes to soil moisture availability in the rooting zone as the mechanistic link between precipitation change and impacts to rates of both Rs (Fig. 4b) and ANPP (Fig. 3b). For Rs, model selection yielded the interaction of soil moisture and temperature as the most parsimonious model for drivers of soil respiration, yet it is likely that the effect of soil temperature on Rs depended on soil moisture availability and indirectly on plant cover (Fig. 4; Fig. S). Specifically, we observed a negative relationship between Rs and temperature (Fig. 4d) that is inconsistent with the generally accepted understanding of this relationship (Lloyd and Taylor 1994). However, higher soil temperatures occurred in drought plots (Fig. 4c) likely due to decreased plant cover and greater light penetration to the soil surface (Maes and Steppe 2012; Fig. S1). We posit that this produced an indirect effect that both increased soil temperatures and exacerbated evaporative soil moisture losses, resulting in an apparent negative relationship of soil temperature with both soil moisture and Rs (Fig. 4). We thus suggest that precipitation-induced soil moisture change, rather than temperature, was the dominant driver of Rs responses to precipitation change in our experiment (Fig. 4), which is consistent with previous findings within this grassland (Hoover et al. 2016) and many other terrestrial ecosystems (Liu et al. 2016).

ANPP responses to precipitation were similarly driven by soil moisture impacts (Fig. 3b), which underscores the critical links between precipitation, soil moisture availability, and primary production within this grassland (Briggs and Knapp 1995; Knapp 2002). This link emphasizes the need for current ecosystem models to represent relationships between precipitation, soil moisture, and primary production, as precipitation alone can be a crude proxy for plant water availability. We further suggest a key yet still underappreciated consequence of these linkages is that predicted changes to within-year rainfall patterns (IPCC 2012), through altering soil moisture dynamics (Knapp et al. 2008), may also impact future precipitation–production relationships in concert with changing precipitation amounts (Nippert et al. 2006).

For example, previous manipulations imposing larger yet fewer rainfall events—without attendant changes to total precipitation amount—were documented to reduce both ANPP and Rs within this grassland (Knapp 2002). Increased precipitation variability reducing carbon exchange processes within this grassland has been attributed reductions in mean soil moisture availability in the rooting zone (Fay et al. 2003; Knapp et al. 2008). Yet by extension, this reduction in mean soil moisture impacted the temporal ANPP–precipitation relationship of this grassland, independent of precipitation amounts. Although our experiment was not designed to manipulate intra-annual precipitation variability, it is evident that concurrent changes to within-year rainfall regimes can impact ANPP–precipitation relationships by way of altering soil moisture. Our results are thus consistent with previous findings that impacts soil moisture may be the underlying determinant for ANPP and Rs responses to changes in precipitation regimes and extremes across many terrestrial ecosystems (Heisler-White et al. 2008; Knapp et al. 2008; Gherardi and Sala 2015; Liu et al. 2016).

Implications for relationships between precipitation and ecosystem processes

Our results do not support our hypothesis of a unimodal response of Rs to soil moisture, as predicted by theory (Moyano et al. 2013). However, sensitivity of Rs can vary considerably across ecosystems of differing climatic backgrounds (Hawkes et al. 2017), suggesting that the patterns observed within this grassland may indeed differ within more historically xeric or mesic ecosystems. Our results may also reflect the temporal resolution of our measurements (Liu et al. 2016). Specifically, because our measurements occurred at least 3 days after water additions, Rs may have recovered from—and thus be highly resilient to—declines induced by saturated soil conditions in this grassland. This uncertainty suggests the need for higher temporal resolution measurements of Rs in precipitation manipulation experiments to better elucidate response dynamics. A unimodal relationship of soil respiration with water availability, if it exists, may be highly transient in this grassland, suggesting a preference for a nonlinear, saturating relationship in benchmarking current Rs–moisture relationships within this ecosystem.

In contrast to our hypothesis, we were unable to detect strong evidence for deviation from a linear ANPP–precipitation relationship towards a nonlinear saturating relationship, as observed previously (Hsu et al. 2012) and predicted by theory (Knapp et al. 2017a). We think this result may be due to two key limitations of our experimental design, which at the same time we suggest presents new avenues of research that can be addressed through regression experiment approaches. First, each precipitation level in our gradient represents a single growing season, and thus it is likely that nonlinear responses of ANPP within our study were difficult to detect due to the lack of time to produce cumulative effects on ecosystem properties. Biotic (physiological adjustment) and abiotic (soil moisture carryover) buffering mechanisms under drought, and high sensitivity of herbaceous vegetation to soil moisture surpluses (Knapp and Smith 2001) may further render a tendency toward a linear relationship for single-year changes in precipitation within this annually burned grassland. A potential pathway towards nonlinearity in ANPP responses to precipitation may thus be through multi-year periods of extremity that produce cumulative impacts on ecosystem properties. For example, Hoover et al. (2014) observed reductions in ANPP to be threefold greater in the second year of an extreme drought, suggesting potential factors such as soil moisture or nonstructural carbohydrate depletion carried over to further constrain ANPP in the second year.

Second, while our gradient included multiple levels of extremes, such extremes were still restricted to percentiles derived from the distribution of the long-term record of the region, and thus were within the bounds of historical climate variability. This suggests that for single-year extremes, magnitudes of extremity with no historical analog must occur (Kayler et al. 2015) to push the system beyond an ecological response threshold (Gutschick and BassiriRad 2003; Smith 2011), and thus towards a nonlinear ANPP–precipitation relationship (Knapp et al. 2017a). As a consequence, investigations concerning how characteristics of precipitation and more generally climate extremes such as the timing, magnitude, and duration, will independently and interactively impact ecosystem processes will be invaluable in projecting responses to extremes.

Conclusions

In summary, we found strong evidence for differential sensitivities of carbon inputs (ANPP) versus losses (Rs) to changes in both precipitation amount and extremity at the site level. Despite our hypothesis of greater ANPP sensitivity to drought not being supported, the higher relative sensitivity of ANPP to wet precipitation extremes within this grassland that we observed is consistent with previous observations (Knapp and Smith 2001; Wilcox et al. 2017). Coupling ANPP responses with the symmetrical sensitivity of Rs to extremes suggests the potential for large carbon gains during wet years within this grassland. The potential for large carbon gains during wet years is further supported when considering the observed linear versus saturating responses of ANPP and Rs to water availability, respectively. Our results additionally suggest that current linear ANPP–precipitation relationships (up to 1000 mm) with mesic, sub-humid grasslands may hold as appropriate benchmarks for ecosystem models, even after incorporating responses to precipitation extremes. By contrast, saturating relationships should be considered as appropriate benchmarks of Rs–precipitation relationships, with changes to soil moisture in the rooting zone as the underlying mechanism driving responses of both ANPP and Rs to changes in precipitation.