Introduction

Soil δ15N values tend to decrease with increasing mean annual precipitation (MAP) and decreasing mean annual temperature (MAT) across a broad range of climate and ecosystem types (Amundson et al. 2003). To some extent, this variation in soil δ15N values is associated with vegetation inputs, given that foliar δ15N values range over 35‰ across plants globally (Craine et al. 2009). Soil δ15N values, however, increase with decreasing soil organic C as global soil organic C concentrations also decline with increasing MAT and decreasing MAP (Craine et al. 2015b). As a consequence, the dependence of soil δ15N on MAP and MAT has been ascribed to this association of soil C with environmental variables and the consequences of these for microbial transformation of both C and N. Furthermore, soils with greater clay concentrations often have higher soil δ15N values. The dependence of soil δ15N on soil C and clay is through fractionation associated with decomposition of soil organic matter that might at least partially be due to better water retention by clay, further linking it with environmental variables (SOM; Craine et al. 2015b).

Like soil δ15N, global patterns of soil δ13C values are correlated with MAP and MAT (Lu et al. 2004), but also with soil texture (Sollins et al. 2009). The largest influence on soil δ13C values, however, is the δ13C value of the input of C to the soil organic carbon (SOC) pool, which is either directly or indirectly derived from primary productivity (Kuzyakov and Domanski 2000). As a consequence, soil δ13C has been used as a proxy for historical vegetation shifts in the distribution of C3 and C4 vegetation (Swap et al. 2004; Gillson et al. 2004; Kuzyakov et al. 2006; Gillson 2015). Despite these clear geographic differences, changes in soil δ13C with depth do not necessarily reflect historic changes in the relative inputs of C3 and C4 vegetation. Turnover processes during soil development also contribute to changes in soil δ13C (Cerling 1984; Balesdent et al. 1993; Qiao et al. 2014) with more decomposed SOC having higher δ13C values (Boström et al. 2007). Thus both soil δ15N and δ13C values are, at least partially, determined by soil processes (i.e. decomposition and mineralization via microbial processing of OM), which may link the patterns of fractionation of these isotopes in the soil. If soil N and C isotope patterns are at least partially linked through common soil processes (i.e. decomposition and mineralization), then we may expect coordinated changes in δ15N and δ13C values with depth through a soil profile.

The δ13C values of SOM through soil profiles commonly increase by 1–3‰ as depth increases below 0.2 m relative to that of the surface litter layer (Chen et al. 2005; Boström et al. 2007). The enrichment of 13C with depth has been shown to occur in tropical, temperate and boreal systems (Hobbie and Ouimette 2009). Although atmospheric δ13CO2 has declined by 1.5‰ over the past 100 years, this has been shown to contribute only marginally to the enrichment of soil δ13C with depth (Ehleringer et al. 2000; Esmeijer-Liu et al. 2012). At least four hypotheses have been proposed for C isotope fractionation through soil profiles. Firstly, kinetic discrimination against 13C during respiration may result from microorganisms preferentially respiring CO2 that is 13C–depleted relative to the substrate, resulting in 13C enrichment of the remaining SOC (Ågren et al. 1996). Although some studies show large 13C depletion of the CO2 formed (e.g. Fernandez et al. 2003), others show no or only minor isotopic fractionation (e.g. Ekblad and Högberg 2000). Secondly, microorganisms are 13C–enriched by 2 to 4‰ compared to plant material (Hobbie et al. 1999) and thus influence SOM, resulting in decreasing C:N ratios with soil depth (Wallander et al. 2003), and compound-specific shifts in soil organic matter to higher δ13C values in products of microbial origin (Huang et al. 1996; Ehleringer et al. 2000). Thirdly, variable mobility (e.g. fulvic acids; Heil et al. 2000) and sorption of isotopes of dissolved organic C on soil particulates (especially clay) may contribute to soil δ13C profiles (Craine et al. 2015b), although some authors have questioned the significance of these mechanism (Boström et al. 2007). Finally, although preferential utilization of 13C–depeleted compounds has been suggested (Boström et al. 2007), the more recalcitrant C fractions of plant biomass (e.g. lignin, lipids and cellulose) that accumulate at depth (Rovira and Vallejo 2002) are 13C–depleted relative to the whole plant (Wilson and Grinsted 1977), and thus cannot contribute to increased 13C–enrichment with depth (Wynn et al. 2006). Apart from this, some, or all, of these processes may thus contribute to determining soil δ13C values to variable extents in different ecological contexts.

As with δ13C, δ15N values usually increases with soil depth, although occasionally maximum δ15N is evident at an intermediate depth possibly as a result of increased volatilization in this soil zone (Hobbie and Ouimette 2009) followed by a subsequent decline at greater depths. The degree of enrichment that δ15N undergoes through a soil profile can have a much broader range than δ13C. In arid and semi-arid systems where soil pH is high, surface δ15N values can be elevated by as much as 7‰ relative to deeper soils (Pataki et al. 2008). There are six potentially important mechanisms that influence δ15N values within soil profiles. Firstly, depletion of 15N by mycorrhizal fungi and transfer of that 15N–depleted N to plants (Hobbie and Ouimette 2009) results in the accumulation of 15N–enriched N derived from mycorrhizal fungi (Hogberg 1997; Hobbie and Ouimette 2009). Secondly, depletion of 15N through enzymatic hydrolysis (Silfer et al. 1992), ammonification, nitrification, or denitrification and the associated fractionation during gaseous loss of 15N–depleted N-containing gas or leaching loss of 15N–depleted NO3 − and the preferential utilization of 14N by plants, drives soil δ15N values up (Handley and Raven 1992; Austin and Vitousek 1998). Thirdly, mixing of soil N among different soil layers through bioturbation (Gabet et al. 2003) and trophic fractionation (i.e. faunal processes; Ponsard and Arditi 2000) could alter soil δ15N profiles. Fourthly, soil texture (i.e. clay) may moderate 14N gaseous loss pathways and/or the differential retention of 15N–enriched SOM (Craine et al. 2015b). Fifthly, preferential microbial utilization of 14N compounds could contribute to accumulation of 15N–enriched compounds deeper in the soil (Boström et al. 2007). Finally, N deposition has been shown to decrease δ15N values of soils because deposited N is typically depleted in 15N, although this effect is relatively small (Liu et al. 2017; Esmeijer-Liu et al. 2012).

SOM decomposition is thus common to both δ13C and δ15N fractionation in soil. At the global scale, climate influences decomposition through both temperature and moisture (Gholz et al. 2000). The SOM composition and nutrient concentrations (especially N) also strongly affect decomposition (Parton et al. 2007). Although most SOM is derived from plants, only a small fraction of the yearly litter and root inputs are incorporated into the stable organic matter pool, most of it after repeated processing by soil microbes (Lerch et al. 2011). SOM transport through soils is generally downward through advection and soil development, and thus the effects of decomposition on soil δ13C and δ15N values are more noticeable deeper in the soil profile. With increasing depth, SOM is more highly processed by microbes (Trumbore 2009) with lower C:N ratios (Marin-Spiotta et al. 2014) and increasing δ13C and δ15N values (Heil et al. 2000; Billings and Richter 2006). This change in δ13C and δ15N is often modelled as “Rayleigh distillation”, which predicts soil δ13C and/or δ15N values based on the soil [C]/[N] in order to account for microbial isotopic enrichment of SOM during decomposition (Mariotti et al. 1981; Baisden et al. 2002; Wynn et al. 2005; Fischer et al. 2008). This enrichment results from the kinetic fractionation during microbial processing (Dijkstra et al. 2006) with subsequent stabilization of products by fine mineral particles in soils (Wynn et al. 2006). This Rayleigh distillation model, however, only pertains to closed systems, potentially ignoring continuous inputs (Fry 2006) that do occur in soils.

Although a number of different factors influence the isotopic fractionation of C and N isotopes, δ13C and δ15N values both increase with soil depth and commonly follow similar trajectories. We hypothesized that changes in soil δ13C and δ15N values are coordinated, possibly through decomposition-related processes, and that the scale of decomposition related changes in δ13C may confound interpretation of soil δ13C as indicative of prior C3 or C4 vegetation. Although the initial isotope composition of the organic matter is indisputably important, subsequent soil fractionation may result in δ13C and δ15N following similar trajectories in space and time. We therefore predict that changes in δ13C and δ15N values correspond with each other both locally through soil depths at a site and globally due to the extent of decomposition and other soil processing. In order to test these predictions, we compiled data from soil depth profiles from sixteen widely distributed sites and also conducted an analysis of global δ13C and δ15N variations in surface soils in order to determine relationships between soil isotopes with climate and soil properties.

Methods

Data sources

Data for soil δ13C and δ15N values were acquired from literature and by contacting individual researchers known to have collected soil isotope data in the past. Soil depth-profile data included δ13C and δ15N for mineral soils at multiple depths at a single site. A second independent dataset included both mineral soil δ13C and δ15N values at a single depth at a number of geographic locations. For each site, climate data were taken from the original source and also, using the geographic coordinates, from the 50-year climatic means (1950–2000) obtained from www.worldclim.org (accessed Sep 2014) at ca. 1 km2 resolution. Variables included were mean annual temperature (MAT), mean annual precipitation (MAP) and 17 other derived climatic variables (Supp. Table 1).

Potential evaporation (PET) was obtained from Trabucco and Zomer (CGIAR Consortium for Spatial Information, 2009. Accessed: http://www.csi.cgiar.org) in which PET was modelled using the method of Hargreaves et al. (1985) with data from Hijmans et al. (2015) and verified by comparison with separate data sources. From the climatic data, the monthly PET was subtracted from monthly precipitation to obtain an index of water availability (P–PET) and averaged to obtain the annual average. Normalized difference vegetation index (NDVI) data was obtained from eMODIS TERRA (US Geological Survey Earth Resources Observation and Science Center), which is corrected for molecular scattering, ozone absorption and aerosols. The NDVI data spanned between 19/12/2009 to 18/12/2012 and was at a spatial resolution of 250 m. The data was averaged to obtain monthly and annual average values using the “raster” (Hijmans et al. 2015) and “RCurl” (Lang and Lang 2016) packages in R.

The fraction of the vegetation with C4 photosynthesis was obtained from Berry et al. (2009) in which the percentage of vegetation within each one degree by one degree grid cell of the land surface which possesses the C4 photosynthetic pathway was determined using ‘C4 climate map’ from Collatz et al. (1998), ‘Continuous fields of vegetation characteristics’ from DeFries et al. (2000) as well as ‘Cropland fraction distribution’ from Ramankutty and Foley (1998). Where necessary, the component fields were re-sampled to bring them to a common one degree by one-degree spatial resolution.

The “SoilGrids1km” global soil data product (Hengl et al. 2014), which has mean soil information at 1 km resolution for six soil depths to 1.5 m deep (ISRIC – World Soil Information 2013), was averaged across the full depth by depth weighted-averaging. The environmental data included in the models is shown in Supp. Table 1.

Soil depth data

Data for 9 sites, which include 4 sites in Africa (Paulshoek, Pretoriuskop, Satara, Hluhluwe) and sites in Alaska, France, Sweden, New South Wales and the Amazon in Brazil were compiled from a number of publications (Table 1). These made up a total of 16 different sampling groups within distinct vegetation types and included data for 79 soil profiles at multiple depths. As most sites were represented by repeated sampling of different vegetation types, the average value of the N and C isotopes at each depth for each vegetation type, as well as the confidence intervals, were determined for each site. As the ranges of δ13C and δ15N values through soil profiles were different in magnitude, the actual measured values were scaled using the “scale” function in R (z-transformation). This allowed both the N and C isotope patterns through the soil profiles to be plotted on the same set of axes for comparison using the ‘ggplot2’ package (Wickham 2009) in R. A Pearson correlation test was then performed on the scaled data. This correlation was then treated as a derived variable. As one of the locations, Hluhluwe, consisted of a number of different vegetation types, each vegetation type at the site was plotted separately rather than averaging across the site.

Table 1 List of sites used in determining the correlation between soil δ13C and δ15N values through soil profiles. Variables included are mean annual temperature (MAT), mean annual precipitation (MAP), δ15N, δ13C, the dominant vegetation at the site, the Pearson correlation coefficients between δ13C and δ15N through soil profiles with significance values (bold where significant, p < 0.05). Values for δ13C and δ15N are include the 5 percentiles, (means) and 95 percentiles of the soil profile data
Full size table

Global analysis of surface soil

In order to determine the main global correlates of soil δ15N values, the dataset from Craine et al. (2015b), which included soil and climatic data for sites around the globe, was re-analysed. Records that did not include a depth or mineral soil components were removed leaving a total of 5447 sites for the analysis. As the δ13C and δ15N values were from single depths only, the dataset was used to determine the global correlation of soil δ13C and other variables with δ15N.

Boosted regression tree analyses

Boosted regression tree models were used to determine how differences in soil and environmental conditions influence the correlation between δ13C and δ15N values for soil depth-profiles, as well as the main drivers of δ15N at the global scale. Boosted regression tree analysis is a form of non-linear modelling that uses machine learning (Elith et al. 2008). The modelling entails decision trees splitting the data into two homogenous groups, a process repeated many times (boosting) so as to improve the prediction of the response variable. Models are parameterized by adjusting their learning rates, tree complexity and bag fraction (Elith et al. 2008). We used a cross-validation procedure to identify the optimal number of trees and tree size for the model, and to guard against over-fitting (Hastie et al. 2001). Initially, the data set was randomly divided into 10 mutually exclusive subsets of equal size, 9 of which were used as a training set to create the boosted tree while the remainder was used as a test set to determine the predictive accuracy of the model. The data in the training sets were fitted using trees of different sizes (range = 2 to 10) by incrementally adding trees in sets of 50. For each combination of tree size and number of trees, the predictive accuracy of the model was determined by comparing values in the test set with those predicted by the model. This procedure was repeated 10 times so that all groups were used as cross-validation groups, and the mean predictive error calculated across all subsets for each level of complexity. The combination of tree size and tree number that produced the lowest predictive error was chosen for all subsequent analyses. Performance was evaluated by expressing the predictive deviance of 10-fold cross validation as a percentage of the null deviance.

Two different models were used, either to explain the correlation of δ13C and δ15N values at the local scale across soil depths (BRTlocal), or to explain the value of δ15N at the global scale for a single soil depth (BRTglobal). The climatic and soil variables listed in Supp. Table 1 were used as the predictor variables. The ‘select07’ function (Dormann et al. 2013) in R, was used to identify collinear predictors. In cases where the predictor variables were found to be strongly collinear with each other, the variable with either the strongest correlation with the response variable, or the most biologically relevant, was retained. Following an initial run (learning rate = 0.01, tree complexity = 5, bagging fraction = 0.5), a simplification procedure was implemented (Elith et al. 2008) to eliminate variables with low influence (such as NDVI and PET). Both models were run ten times using the libraries ‘gbm’ (Ridgeway et al. 2013) and ‘dismo’ (Hijmans and van Etten 2014) packages in R. Model outputs were used to ascertain the relative influence and relationship of each predictor with the correlation between δ13C and δ15N at the local scale or δ15N at the global scale.

To account for C3 and C4 vegetation input into the SOM pool, global soil δ13C values were analyzed for bimodality using libraries ‘diptest’ (Maechler 2015) in R and cutoffs were calculated using the ‘mixtools’ (Benaglia et al. 2009). δ13C for C3 and C4 were then treated as separate sets of data on which BRT modeling for global δ15N values were independently reanalyzed.

Results

Isotopic variation with soil depth

For 11 out of 16 sampling groups analyzed, the variation in average soil δ13C and δ15N values with depth were significantly positively correlated with each other (Fig. 1, Table 1). For many of these sites, both δ13C and δ15N values increased with depth, with the majority of the increase occurring in the upper 10–20 cm of the profile. The range of variability for both isotopes was ca. 2–8‰ through the soil profiles and this range was independent of the average δ13C and δ15N signature for the sites (Table 1). Within the relatively small geographic area of the Hluhluwe Nature reserve, the significant positive correlations between δ13C and δ15N values were independent of vegetation types comprising forest, grassland, savanna and thicket sites. Across all of these distinct vegetation types, δ13C and δ15N values increased similarly with depth (Fig. 2, Table 1). For these sites the range of variability for both isotopes was also ca. 2–8‰ with the majority of the increase in δ13C and δ15N values occurring within the upper ca. 20 cm of the soil. Although most sites had significant positive correlations between δ13C and δ15N, for 5 of the 16 sampling groups, changes in average soil δ13C and δ15N values through the soil profiles were either not significantly associated or negatively correlated with each other (Fig. 3, Table 1). For these sites δ13C and δ15N values also increased with depth, with the exception of the Paulshoek site in which δ15N initially increased before subsequently decreasing below ca. 10 cm. These sites also had a wider range of δ13C and δ15N values than those for which there were significant correlations between δ13C and δ15N (Figs. 1, and 2).

Fig. 1
figure 1

Variation with soil depth of δ13C and δ15N values for sites in which δ13C and δ15N are significantly correlated with each other (Table 1). The data was averaged for each depth and the confidence interval is represented by the coloured bands. The δ13C and δ15N data were independently centred on 0 so as to allow comparison of the variation of these within a site and thus the range of the data corresponds to that of the original data. Sites designated OC and UC are from open-canopy and under-canopy, respectively

Full size image
Fig. 2
figure 2

Variation with soil depth of δ13C and δ15N values for sites in which the dominant vegetation types differ. The data was averaged for each depth and the confidence interval represented by the coloured bands. The δ13C and δ15N data were independently centred on 0 so as to allow comparison of the variation of these within a site and thus the range of the data corresponds to the original data

Full size image
Fig. 3
figure 3

Variation with soil depth of δ13C and δ15N values for sites in which δ13C and δ15N are poorly correlated with each other. The data was averaged for each depth and the confidence interval represented by the coloured bands calculated from the standard error. The δ13C and δ15N data were independently centred on 0 so as to allow comparison of the variation of theses within a site. The range of the data corresponds to the original data

Full size image

BRT analysis of the correlation between δ13C and δ15N values ranked CEC, mean diurnal temperature range, bulk density, MAT, clay and MAP as the top predictors (Fig. 4a), explaining 38% of the variance in the correlation between δ13C and δ15N. Partial dependency plots, which show the effect of a variable on the response after accounting for the average effects of all other variables in the model, of the BRT analysis of the soil profile correlations between δ13C and δ15N values (Fig. 5), showed that this was strongest at sites with CEC < 20 cmol kg−1 and a mean diurnal temperature range < 13°C. Sites with bulk density above 1400 kg m3 had strong correlation between soil δ13C and δ15N values. The influence of clay concentration on the correlation between δ13C and δ15N values was generally high. A number of sites with clay concentrations between 30 and 35%, however, had a relative low influence of clay on the correlation. These sites were arid, receiving <500 mm mean annual precipitation and had a relatively poor correlation compared to mesic sites (i.e. between 500 and 1000 mm) with a moderate influence in hydric sites (>1000 mm). The correlation between δ13C and δ15N values was stronger at sites with MAT >19°C (Supp. Fig. 4f).

Fig. 4
figure 4

Relative influence of variables in determining the correlation between global soil δ13C and δ15N as determined by BRT analysis (a) as well as the relative influence of variables in determining the global soil δ15N as determined by BRT analysis (b). Values are the mean ± SE of 10 runs of each model. Error bars represent standard error

Full size image
Fig. 5
figure 5

The global variation in soil δ13C and δ15N. The color of the points represents the site averages of δ13C and δ15N values standardized and centered to range between −1 and 1. Background fill colour represents mean annual temperature

Full size image

Global geographic variation

Globally, soil δ15N values of surface soils were significantly positively correlated with δ13C, MAT and the prevalence of C4 photosynthetic vegetation and negatively correlated with CEC and diurnal T range (Table 2). Geospatial variation in global δ13C and δ15N values that were spatially averaged over 0.1° corresponded relatively well with each other at high latitudes (> 50°) where both δ13C and δ15N values were more negative compared to sites located nearer the equator (Fig. 5). Sites in which δ13C values were relatively high (Fig. 5) were from more arid regions such as Southern Africa, Australia and North America and in which C4 grass communities exist (Fig. 6).

Table 2 Bivariate ranged major axis (RMA) analysis results of top six predictors of global soil δ15N with correlation coefficients (r) shown with p-values (bold where significant). All variables used in the prediction of global soil δ15N are shown in Supp. Table 1
Full size table
Fig. 6
figure 6

Bivariate analysis of the top six predictors of global soil δ15N against global soil δ15N. Lines indicate linear model function

Full size image

BRT analysis of global soil δ15N values (BRTglobal) ranked MAT, δ13C, CEC, C4, diurnal range and MAP as the top predictors of soil δ15N (Fig. 4b), which explained 62% of the variation in δ15N values. The partial dependency plots for the BRTglobal (Supp. Fig. 5) showed that as MAT increased, δ15N values also increased. Sites with δ13C values below ca. -30‰ had low δ15N values, which increased rapidly with increased δ13C values up until ca. -20‰, above which changes in δ15N values were relatively small. Therefore, much of the change in δ15N values associated with δ13C values occurred in a range of δ13C values considered to be characteristic of C3 dominated sites (Supp. Fig 2). Sites with CEC values >10 cmol kg−1 had relatively low soil δ15N values. δ15N values were also low for sites with <75% C4 vegetation. Soil δ15N values were reduced with increases in mean diurnal temperature range and generally with increased MAP (Supp. Fig. 5f).

Global δ15N values predicted from the full BRTglobal model, including both C3 and C4 sites, were strongly correlated with observed global δ15N values (Supp. Fig. 1). There was, however, a degree of under-prediction of δ15N values at low observed δ15N values and over-prediction at high observed δ15N values. Global soil δ13C values were bimodal with two ranges of δ13C values having peaks at −26.36‰ and −17.58‰, indicating that there were a number of sites dominated by either predominantly C3 or C4 plants (Supp. Fig. 2). BRT’s predicting global soil δ15N based on a subset of sites that were predominantly C3 dominated ranked MAT, δ13C, CEC, bulk density, diurnal T range and MAP as top predictors (Supp. Table 2). The BRT developed for C4 dominated sites ranked MAT, CEC, bulk density, MAP, diurnal range and δ13C as top predictors. Although soil δ13C was found to be a strong predictor of δ15N for C3 sites, it was a weak predictor in C4 dominated sites.

Discussion

This study suggests that either common or coordinated processes contribute to fractionation of soil C and N isotopes. The link between soil δ13C and δ15N values may inform understanding of these processes due to this coordination of soil processes determining both C and N isotope fractionation. Our results suggest that although the initial isotope composition of the organic matter is indisputably important, subsequent fractionation via soil processes, such as decomposition and related processes, may result in correlations between δ13C and δ15N values in geographic space and commonly following similar trajectories with soil depth. More positive δ13C and δ15N values with soil depth (Fig. 1) must result from increasing fractionation or more prolonged fractionation in deeper soils relative to shallower soils.

The importance of the vegetation characteristics in determining C isotopic composition is apparent from the bimodal distribution of soil δ13C values associated with C3 (−22‰ to −32‰; Troughton 1979) and C4 (−9.2‰ to −19.3‰; Hattersley 1982) vegetation (Fig. S2) whereas the variation in δ13C within the C3 and C4 groupings is caused by climatic and geographical factors (Damesin et al. 1997). Likewise, global variation in soil δ15N values (Fig. 5) is associated with variation in foliar δ15N that varies with MAP, MAT, N availability, foliar N concentration, species composition and with the degree of N2 fixation (Craine et al. 2009). Organic matter enters soils in a diversity of ways and this influences the initial isotopic signature of soil C and N (Eissfeller et al. 2013). The majority of SOM, however, enters the soil as plant-derived detritus, where it is utilized by soil microbes (Berg and McClaugherty 2008) and decomposer fauna (Hättenschwiler and Gasser 2005). Consequently, the isotopic values of the dominant vegetation and the variation in δ13C and δ15N values, both between and within species (Damesin et al. 1997; Craine et al. 2015a), strongly influence SOM isotopic composition.

Unlike for C, however, there are also strong ecosystem feedbacks between soil and vegetation N in determining ecosystem δ15N values, because soil δ15N also partially determines plant δ15N. Despite this dependence of SOM isotopic composition on that of OM and vegetation, the variations in δ13C (range: −27.8 to −12.4‰) and δ15N (range: −0.1 to 10.1‰) with depth in soil profiles were often strongly correlated with each other (Table 1). Likewise, geospatial variation in global δ13C and δ15N values also corresponded relatively well across a wide range of climates and biomes (Fig. 5). For example, C3 and C4 dominated sites showed similar patterns of δ13C and δ15N enrichment through soil profiles (Fig. 2), although the range of values was smaller with C4 vegetation.

The correspondence between the increases of δ13C and δ15N values with depth is probably through processing of SOM, which is further supported by the most influential predictors in the BRT model for the correlation between δ13C and δ15N values through soil profiles (Fig. 4a), which themselves are related to microbial activity. Furthermore, soil δ13C values were also strong determinants of δ15N globally (regardless of soil and ecosystem type) while the remaining top predictors of δ13C could be related to SOM decomposition (Fig. 4b). Processing of SOM is determined by characteristics of the SOM, such as the C and N composition (Fernandez et al. 2003), as well as by environmental factors including soil temperature, moisture and aeration (Gholz et al. 2000; Zhang et al. 2008). The reason for the positive correlation between MAT and both δ15N and δ13C values could therefore be due to microbial activity increasing with increasing temperature. Mean diurnal temperature range (e.g. Li et al. 2011), CEC and soil fertility (Sikora 2013) may also be linked to SOM decomposition through soil microbial processes. Although favorable moisture conditions stimulate decomposer communities (Cotrufo et al. 2013), MAP was not significantly correlated with either δ13C or δ15N values at the global scale (Fig. 4b, Table 2). This is likely because many ecosystem properties depend on MAP obscuring clear relationships. For example, Craine et al. (2015a) related variation in global soil δ15N to variation in clay concentrations. Further, there is the possibility that the limited range in MAP at the regional scale can obscure relationships between soil δ15N and MAP as the increase in soil δ15N with increasing MAP at the regional scale often breaks down at broader scales (Amundson et al. 2003; Austin and Vitousek 1998).

Despite strong global geographic correspondence between δ13C and δ15N and correspondence over soil depth (11 of 16 sites), some sites had non-significant (New South Wales, France, Sweden) or negative (Paulshoek) correlations between δ13C and δ15N (Fig. 3, Table 1). These sites indicate the complexity to the relationship between soil δ13C and δ15N, and dependence on other factors. For example, the New South Wales sites had a large proportion of N2-fixing microbes in the surface soil (Macdonald et al. 2015) resulting in δ15N being close to 0‰. The non-significant Swedish and French sites were both associated with plantations (Boström et al. 2007; Zeller et al. 2007), whereas a corresponding natural site in France showed a significant relationship (Fig. 1). Paulshoek exhibited a maximum soil δ15N value at intermediated depths, which is indicative of N-loss during nitrification and denitrification (Hobbie and Ouimette 2009). This is not surprising as Paulshoek is arid with high soil temperatures and sporadic rainfall (Table 1) and these conditions increase nitrification/denitrification rates (Craine et al. 2015b). Thus despite the general global relationship between δ13C and δ15N, this correspondence does vary depending on local biotic, disturbance and environmental influences.

As a consequence of a link between soil δ13C and δ15N, interpretation of soil δ13C values as indicators of historical vegetation assemblages is complicated by the role of soil processes in determining soil δ13C values, as also shown by Wynn et al. 2005. The ranges of δ13C values with depth are commonly large (up to 11.0 ‰, Supp. Fig. 3) which overlaps the range of values commonly associated with vegetation change. For example, δ13C values between −16 and −20‰ have been used to indicate mixed C3 and C4 vegetation and > −16% to indicate C4 dominance (Gillson 2015). From our study, however, whilst the minimum δ13C values of soils with C3 and C4 vegetation reflect the isotopic signature of the vegetation inputs, the maximum δ13C values are indistinguishable. Since the maximum δ13C values of soils supporting C3 vegetation overlap with the minimum δ13C values of C4 vegetation, interpretation of intermediate δ13C values (i.e. < ca. -15 ‰) as indicating historical vegetation characteristics should be approached with caution. Furthermore, in order to demonstrate that ancient δ13C SOC values are indeed representative of ancient vegetation assemblages in samples of deep SOC, one must establish that the fraction of SOC remaining in the sample is very close to the original maximum concentration during soil formation and that fractionation has not been great (Wynn et al. 2006). This is because Rayleigh distillation and mixing processes vary with environmental and soil properties, with particularly strong effects associated with fine mineral particles (i.e. clay) in fine grained soils (Krull and Skjemstad 2003; Wynn et al. 2005) and should not be assumed to be constant everywhere.