Introduction

By burning fossil fuel and applying artificial fertilizer to cropland, humans have increased atmospheric N deposition three- to five-fold over the last century (IPCC 2007; Bouwman et al. 2013; Lamarque et al. 2013). Global N deposition rates are projected to increase by 2.5 times or more by the end of this century (Reay et al. 2008), and much of this N will be deposited in terrestrial ecosystems. Because plant growth is typically limited by N availability (LeBauer and Treseder 2008; Bai et al. 2010; Wright 2019) and N limits the CO2 fertilization on plant biomass (Terrer et al. 2019), increasing N deposition rates are widely expected to stimulate ecosystem C storage (Reay et al. 2008; Janssens et al. 2010; Chen et al. 2018). However, it is still unclear to what extent N additions affect soil C stocks. Soils store about twice as much C as the atmosphere and form a large natural source of CO2 (Eswaran et al. 1993). Thus, soil C responses to N enrichment could play a key role in determining future trajectories of atmospheric CO2 concentrations (Dijkstra et al. 2004; Loisel et al. 2019).

The size of the soil C pool is determined by the balance between soil C input – mostly from plant litter and roots (e.g., exudates and allocation to mycorrhizal fungi) – and soil C output through the decomposition of soil organic matter (Trumbore 1997; Jastrow et al. 2007). By stimulating plant growth and litter production, N input from anthropogenic sources can increase soil C stocks through the enhancement of C inputs (e.g., Gong et al. 2012; Tian and Niu 2015). However, in some cases N enrichment stimulates soil C input without increasing soil C stocks (Mack et al. 2004; Allison et al. 2010) and vice versa (Pregitzer et al. 2008), suggesting that N-induced changes in decomposition (i.e., soil C output) affect soil C storage as well (Janssens et al. 2010). Indeed, N additions can decrease the activity of lignin-modifying enzymes (Chen et al. 2018) and N-induced decreases in soil pH can decrease decomposition rates (e.g. Zhou et al. 2017); both these responses would stimulate soil C storage.

Because most N enrichment experiments do not directly measure the fate of newly added plant litter vs. native soil organic matter, they provide limited mechanistic insight in the processes underlying soil C sequestration (Cardon et al. 2001). However, the dynamics of these two different C pools can be studied through isotopic labeling, in which the isotopic signature of soil C inputs differ from pre-existing soil C (from now on referred to as “old soil C”). This approach enables us to determine the amount of soil C derived from old soil C versus “new soil C” (i.e., cumulative plant inputs since labeling began) (Keith et al. 1986; Balesdent et al. 1987). Similarly, by analysing the isotopic composition of total soil CO2 respiration in labeling experiments, we can quantify the decomposition of old soil C (Rochette et al. 1999).

The response of soil C storage to N addition varies with environmental conditions and between ecosystems. For instance, field experiments suggest that N enrichment generally does not alter soil C stocks in grasslands (Lu et al. 2011), but stimulates soil C sequestration in temperate, tropical, subtropical and boreal forests (Janssens et al. 2010; Cusack et al. 2011; Frey et al. 2014; Maaroufi et al. 2019). Furthermore, N enrichment reduced litter decomposition in N-rich ecosystems, but not in N-limited ecosystems (Chen et al. 2015). Effects of N also depend on litter quality; whereas N addition typically stimulates the decomposition of labile C, it slows down the decomposition of recalcitrant C (Fog 1988; Talbot and Treseder 2012; Chen et al. 2018).

Because N enrichment studies cover a wide range of ecosystem types and environmental conditions, deriving a global response to N enrichment from individual experiments is challenging. Moreover, high spatial variability in soil C stocks and low replication mean that individual experiments often lack the statistical power to detect changes in soil C dynamics (Hungate et al. 1995). A quantitative synthesis of results across multiple studies can overcome both these problems. Thus, we used meta-analysis to synthesize studies that applied isotopic labeling to evaluate the impacts of N enrichment on new and old soil C. By synthesizing these data, we aimed to identify the main factors determining new and old soil C stocks with N enrichment, thereby increasing mechanistic insight into the processes underlying soil C storage. We hypothesized that N addition stimulates soil C storage both by increasing plant growth and new soil C storage, and by reducing the decomposition of old soil C.

Methods

Data collection

We used Web of Science (Thompson Reuters) to search for studies published before March 2020 that employed an isotopic C tracer to study the effect of N enrichment on soil C dynamics. We used the search terms “nitrogen”, “soil AND carbon” and “isotop* OR label*” for article topic.

To be included in our dataset, studies had to meet the following criteria. First, studies had to include at least two N treatments (“control” & “high N”), with N addition rates (kg N ha−1 yr−1) for both treatments clearly indicated. Second, to distinguish “new soil C” and “old soil C”, C labeling should be applied to create a difference in the isotopic composition of plants and soils. There are two ways to achieve this goal: (1) growing plants under isotopically labeled CO2 (i.e., 13CO2 or 14CO2); (2) growing C3 plants on soils that developed under C4 vegetation, or vice versa. C3 plants discriminate more strongly against 13C than C4 plants because of differences in the photosynthetic pathway, thereby creating differences in the 13C isotopic composition of plant biomass and soil organic matter input (Farquhar et al. 1989). Thus, in both approaches, new soil C derived from plants will differ from native (i.e., unlabeled) soil C. Third, we only considered studies that applied continuous labeling, usually from the first leaf emergence to sampling time, to ensure that the total amount of new soil C could be determined (Kuzyakov and Domanski 2000). We included both pot- and field- experiments in our analysis. In total we found 35 studies that met our requirements (Table 1).

Table 1 Overview of N enrichment experiments included in our meta-analysis; responses that were reported in each study are indicated by ‘●’
Full size table

From each study we extracted data on new, old, and total soil C, and old soil C respiration when these were reported, for both control and high N treatments (Data S1-S5). For studies reporting new soil C stocks, we also extracted data on soil C input proxies (root biomass, or yield data if root biomass was unavailable), following the approach of van Groenigen et al. (2017). To avoid pseudoreplication, we only included the most recent observations from each study in our dataset. We tabulated means, standard deviation and the number of replicates for both control and high N treatments. Missing standard deviations were estimated from the mean coefficient of variation across the dataset (e.g., van Groenigen et al. 2017).

Previous studies suggest that plant growth and total soil C storage with N addition depend on ΔN (i.e., the difference in N addition between the control and high N treatments; e.g., Maaroufi et al. 2015), plant type (e.g., Yue et al. 2016), atmospheric CO2 concentration (van Groenigen et al. 2006), soil N availability (Chen et al. 2015) and soil pH (Nottingham et al. 2015). Similarly, several studies suggest that N fertilizer stimulates plant growth more strongly when it is applied in combination with other nutrients (e.g., Crowther et al. 2019; Elser et al. 2007). Thus, we categorized studies based on these experimental conditions. To account for plant type, we made a distinction between studies on woody species and studies on herbaceous species. We made a distinction between studies that added other nutrients together with N, and studies that did not. Atmospheric CO2 concentration, soil pH and ΔN were included in our analysis as continuous factors. Our dataset included only 5 studies on woody species, limiting the representativeness of the overall treatment effects for this category.

We used soil C:N ratios as an indicator of initial soil N availability (e.g., Terrer et al. 2019). Because plant growth responses to N addition show a saturating response (Aber et al. 1998), we also included N addition levels in control treatments as a predictor variable. Some pot studies in our dataset applied N fertilizer homogenously throughout the entire soil column by mixing (e.g., Heath et al. 2005). In these cases, we tabulated N addition levels equivalent to the amount of N added to the 0–20 cm layer. Finally, we tabulated information about labeling type (i.e., C3-C4, 13C, or 14C), study method (i.e., pot vs. field), clay content (%) and experimental duration (i.e., the natural log of the number of days since the isotopic label was introduced in the experiment).

Meta-analysis

For each study in our dataset, we calculated the response of new soil C stocks, old soil C stocks, total soil C stocks, C input proxies and old soil C respiration to N enrichment. We expressed treatment effects as the natural log of the response ratio (lnR), a metric commonly used in meta-analysis (Hedges et al. 1999; Osenberg et al. 1999):

$$\ln \mathrm{R}=\mathit{\ln}\left(\frac{V_h}{V_c}\right)$$

with Vh and Vc as the arithmetic mean values of new soil C, old soil C, total soil C, soil C input, and old soil C respiration in the high N and control treatments, respectively. The variance (var) of lnR was calculated as:

$$\operatorname{var}=\frac{SD_h^2}{N_h{V}_h^2}+\frac{SD_c^2}{N_c{V}_c^2}$$

with Nh and Nc as the replicate numbers for high N and control treatments, respectively.

Our meta-analytical approach was adopted from van Groenigen et al. (2017). Briefly, we analysed our data using a mixed-effects model with the rma.mv function in the R package “metafor” (Viechtbauer 2010). Because several studies contributed more than one effect size (e.g. in multifactorial experiments), we included “study” as a random effect. We weighted lnR by the inverse of its variance. Models were fitted according to Knapp and Hartung (2003); 95% confidence intervals (CI) of treatment effects were based on critical values from a t-distribution. Effect sizes were considered significant if their 95% CI did not include zero. We used a Wald test to statistically evaluate differences in treatment effects among categories. The “glmulti” package was used to identify which of the factors described above best predicted treatment effects of N enrichment, following the same approach as Terrer et al. (2016) and van Groenigen et al. (2017). In short, we analysed our data with all possible models that could be constructed using combinations of the experimental factors. Model selection was based on Akaike Information Criterion corrected for small samples (AICc) as criterion. The relative importance value for a particular predictor was equal to the sum of the Akaike weights (probability that a model is the most plausible model) for the models in which the predictor appears. A cut-off of 0.8 was set to differentiate between important and redundant predictors, so that predictors with relative importance near or less than 0.8 are considered unimportant.

Results

Averaged across the entire dataset, N addition significantly stimulated soil C input proxies (+30.7%, p < 0.001), new soil C stocks (+30.3%, p < 0.001), and total soil C stocks (+6.1%, p < 0.001) (Fig. 1). In contrast, N addition did not significantly affect the respiration of old soil C (+0.2%, p = 0.98) or old soil C stocks (+1.2%, p = 0.60). When we limited our analysis to long-term (i.e., >5 year) studies conducted under field conditions, we found quantitatively similar results for all variables (Fig. S1). Unfortunately, no long-term data were available for old soil C respiration.

Fig. 1
figure 1

Results of a meta-analysis on the responses of new soil C stocks, old soil C stocks, total soil C stocks, soil C input proxies and old soil C respiration to N addition. The number of observations (n) and total number of independent studies included in each analysis are displayed below each bar. Error bars indicate 95% confidence intervals. ** and *** indicate significance at p < 0.01 and p < 0.001, respectively

Full size image

Our model selection approach indicated that “Control N” (i.e., N addition level in control treatments) was the most important predictor for N-induced effects on both new soil C stocks and soil C input proxies (Fig. 2a, b). Among all the potential models for the response of new soil C stocks to N additions (“lnRN”), the model lnRN ~ Control N had the lowest AIC value (Table S1). Of all the potential models for the response of C input to N additions (“lnRI”), the model lnRI ~ Control N was the most parsimonious within 2 AIC units (Table S2). Across the entire dataset, lnRN and lnRI both decreased with increasing control N levels (p < 0.05, Fig. 3a, b). Nitrogen-induced changes in soil C input and new soil C stocks were significantly correlated (p < 0.05; Fig. S2).

Fig. 2
figure 2

Model-averaged importance of the predictors of the N-effect on new soil C stocks (a) and soil C input proxies (b). The importance is based on the sum of Akaike weights derived from model selection using AICc (Akaike’s Information Criteria corrected for small samples). See Fig. 1 for the number of observations and independent studies used in each analysis

Full size image
Fig. 3
figure 3

The relationship between control N levels and treatment effects (lnRN) on new soil C stocks (a) and the relationship between control N levels and treatment effects (lnRI) on soil C input (b). Symbol size represents the weight of each observation in our meta-analysis. See Fig. 1 for the number of observations and independent studies used in each analysis

Full size image

The dataset of N effects on old soil C respiration (lnROR) only contained pot studies on herbaceous species. Therefore, the model selection approach for this dataset did not include plant type or study method as possible predictors. Model selection indicated that “ΔN” was the most important predictor for lnROR (Fig. 4a). The model lnROR ~ ΔN was the most parsimonious model within 2 AIC units (Table S3); lnROR decreased by 0.125 units per 100 kg N ha−1 yr−1 change in ΔN (p < 0.05; Fig. 5).

Fig. 4
figure 4

Model-averaged importance of the predictors of the N-effect on old soil C respiration (a) old soil C stocks (b) and total soil C stocks (c). The importance is based on the sum of Akaike weights derived from model selection using AICc (Akaike’s Information Criteria corrected for small samples). See Fig. 1 for the number of observations and independent studies used in each analysis

Full size image
Fig. 5
figure 5

The relationship between N addition levels (ΔN) and treatment effects on old soil C respiration (lnROR). Symbol size represents the weight of each paired observation in our meta-analysis. The analysis is based on 30 observations, derived from 13 independent studies

Full size image

The response of old soil C stocks to N addition (“lnRO”) was best predicted by Control N, ΔN and CO2 concentration (Fig. 4b). Among all the potential models, the model lnRO ~ Control N + ΔN + CO2 concentration was the most parsimonious model within 2 AIC units (Table S4). Analysing our dataset with this model, the effect of N addition on old soil C stocks increased with ΔN (p < 0.001) and atmospheric CO2 concentrations (p < 0.001), but it decreased with control N levels (p < 0.01; Fig. 6).

Fig. 6
figure 6

Results of a meta-analysis on the response of old soil C stocks to N addition (lnRO), based on a model that includes Control N levels, ΔN and atmospheric CO2 level as moderators. Intercept represents the lnRO for Control N and ΔN at 0 kg N ha−1 yr−1 and atmospheric CO2 level at 400 ppm. The analysis is based on 92 observations, derived from 17 independent studies. Error bars indicate 95% confidence intervals. **,*** indicate significance at p < 0.01 and p < 0.001, respectively

Full size image

The response of total soil C stocks to N addition (“lnRT”) was best explained by atmospheric CO2 concentrations (Fig. 4c). Among all the potential models, “lnRT ~ CO2 concentration” had the lowest AIC value (Table S5). As with old soil C stocks, the effect of N addition increased with atmospheric CO2; lnRT increased by 0.012 units per 100 ppm change in atmospheric CO2 concentrations (95% CI: 0.003–0.021; p < 0.01).

Discussion

New soil C plays a key role in soil C dynamics, since its quantity and quality influence soil C decomposition rates (e.g., De Graaff et al. 2010). The positive correlation between N-effects on new soil C stocks and soil C input proxies suggests that, N-induced changes in new soil C stocks are at least partly driven by changes in soil C input. N-effects on soil C input and new soil C stocks both decreased with increasing control N levels. These findings can probably be explained by saturation of the N enrichment effect (Aber et al. 1998). Adding N to N-rich soils often induces shifts in nutrient limitations of plant growth from N to phosphorus (Vitousek et al. 2010; Penuelas et al. 2013; Deng et al. 2017). Under these conditions, further N addition no longer increases plant growth and may even decrease plant growth when competition dominates plant–microbe relationships (Čapek et al. 2018). Furthermore, high N addition rates tend to decrease soil pH (Tian and Niu 2015), thereby causing a loss of base cations and increasing soluble aluminium concentration, both of which negatively affect plant production (Bowman et al. 2008).

Because N addition promotes plant growth more strongly in N-poor ecosystems (e.g., Chen et al. 2015), we expected N additions to increase new C stocks and plant growth more strongly in experiments with high soil C:N ratios. Yet, treatment effects did not depend on soil C:N ratio for studies in our dataset. One possible explanation for this result is that soil disturbance distorted the relation between soil C:N ratio and soil N availability. Experiments in our analysis inherently involve some level of soil disturbance, such as replacing vegetation (i.e., by using soil that developed under vegetation with a different photosynthetic pathway than that of the experimental vegetation) and transferring soil from the field to pots. Furthermore, some experiments in agricultural systems applied disturbance during the experiment in the form of tillage operations. When soil disturbance breaks up aggregates, physically protected soil organic matter becomes available to microbes, stimulating organic matter decomposition and increasing nutrient availability (e.g., Kristensen et al. 2000). Thus, soil disturbance may have contributed to the relatively high variation in treatment effects on new soil C and soil C input in our dataset.

The central role of control N levels in determining the potential of N-induced new soil C storage has important implications for soils both in agricultural and natural ecosystems. Soil C sequestration in agricultural systems has been widely suggested as a main mechanism to reduce anthropogenic CO2 emissions and slow down climate change (e.g., Minasny et al. 2017). Our results suggest that N additions to agricultural soils that previously received little or no additional N could contribute to these efforts. In contrast, further N additions to agricultural systems that already receive substantial amounts of fertilizer N are unlikely to stimulate new soil C storage. Soil emissions of N2O strongly increase when fertilizer N rates exceed crop N uptake (e.g., van Groenigen et al. 2010), suggesting that further N additions in these systems might even work counterproductive in terms of greenhouse gas mitigation. Less intensively managed ecosystems on the other hand typically receive low amounts of N input, mostly through atmospheric N deposition. Thus, our results suggest that future increases in atmospheric N deposition will lead to net soil C sequestration in these systems.

Although N addition on average did not affect old soil C contents or old soil C respiration, lnRO increased and lnROR decreased with ΔN. We speculate that these results reflect the various mechanisms through which N additions can affect old soil C respiration. N additions can stimulate decomposition of native soil organic matter indirectly by increasing plant growth and labile C inputs which act as substrate for soil microbes (e.g. Paterson et al. 2008). On the other hand, N additions can have direct negative effects on lignin-modifying enzymes (Chen et al. 2018), and N-induced decreases in soil pH may decrease decomposition rates (e.g. Zhou et al. 2017). In the absence of plants, N additions typically decrease microbial respiration (Fog 1988; Janssens et al. 2010), suggesting that negative effects of N additions on soil microbial activity dominate when N rates exceed plant N uptake, i.e. at high ΔN. This interpretation is consistent with several studies indicating that N additions decrease soil respiration more strongly at high N addition rates (e.g. Janssens et al. 2010).

The positive relation between atmospheric CO2 concentration and lnRO corroborates studies suggesting that N enrichment reduces old soil C decomposition under elevated CO2 (e.g. Cardon et al. 2001; Cheng and Johnson 1998). It is also consistent with a recent meta-analysis showing that N additions tend to increase old soil C stocks under elevated CO2 (van Groenigen et al. 2017). This result might be explained by N additions reducing CO2-induced priming of soil organic matter by alleviating N limitation of plant growth. Indeed, several studies show that elevated CO2 stimulates decomposition of old soil organic matter, thereby releasing N to support plant productivity (e.g. Cheng 1999; Langley et al. 2009). CO2-induced priming occurs mostly in N limited ecosystems (Dijkstra et al. 2013; Terrer et al. 2018), possibly explaining why N addition reduces CO2-induced decomposition of old C.

Three limitations of our analysis must be noted. Most importantly, our dataset includes relatively few long-term studies under field conditions. For instance, the longest experiment on woody species lasted 6 years, a relatively short period to measure treatment effects on long-living plants. Unfortunately, the subset of long-term field studies is too small for our model selection approach. However, the few long-term field studies that directly tested the impact of important model predictors largely support our findings. For instance, Wilts et al. (2004) found that in a 29 yr old experiment, average new soil C stocks increased with N addition, and lnRN decreased with increased control N levels. In a study with relatively high control N levels (140 kg N ha−1 yr−1) and high ΔN values (420 kg N ha−1 yr−1), van Groenigen et al. (2003) found that N addition tended to decrease new soil C stocks and increase old soil C stocks. However, individual studies are limited by high spatial variability in soil C stocks and the effects described above were not significant. Clearly, more long-term studies are needed to determine whether the factors affecting N-induced soil C storage change over time. Spatial variability in soil C stocks can be reduced by planting communities on homogenized soils (e.g. Cardon et al. 2001; van Kessel et al. 2000), and statistical sensitivity to detect treatment effects might also be improved by increasing the difference in isotopic signature between newly fixed and old C pools (Ogle and Pendall 2015), or by combining isotopic labeling with physical soil fractionation techniques (e.g., Dijkstra et al. 2004).

Second, our dataset only includes studies evaluating the effect of N enrichment as a single factor. However, terrestrial ecosystems are currently exposed to multiple types of environmental change, and these changes are likely to interact. For instance, rising levels of atmospheric CO2 stimulate plant growth and soil C storage, especially when combined with N additions (van Groenigen et al. 2006; Terrer et al. 2018). Thus, under future atmospheric CO2 concentrations, N saturation of plant growth and soil C storage may occur at higher N rates than under current CO2 levels.

Finally, our analysis does not include any studies conducted in the tropics. Because plant productivity in the tropics is usually limited by both N and P, plants growth responses to N enrichment may saturate at relatively low levels (Wright 2019). Thus, the potential for N-induced C sequestration might be smaller in tropical soils than in temperate soils. Together, these three limitations underline the need for more long-term, multi-factor global change experiments on a wider range of ecosystems to study the fate of future soil C stocks.

In summary, our analysis indicates that the response of soil C dynamics to N addition depends on numerous environmental factors and varies strongly between experiments. However, our results suggest that N enrichment increases new soil C stocks substantially. Nitrogen-induced increases in new soil C sequestration are strongest in ecosystems receiving low amounts of N, indicating that the positive effect of N deposition on soil C storage likely diminishes with continuous N enrichment. Furthermore, we found that decomposition of old soil C decreased at high N addition levels. Thus, we provide isotopic evidence that N addition can stimulate soil C storage both by increasing soil C input and by decreasing decomposition rates. These results suggest that to improve prediction of future soil C storage with N enrichment, changes in soil C input and decomposition should both be considered, along with temporal changes in soil N status.