Soil–climate interactions explain variation in foliar, stem, root and reproductive traits across temperate forests
Abstract
Aim
The individual effects of climate and soil properties on functional trait distributions have become increasingly clear with recent syntheses of large datasets. However, the distribution of traits in a given climate may depend on the fertility of the soil. Our aim was to quantify how soil–climate interactions explain community-level variation in functional traits from every plant organ to improve predictions of plant community responses to environmental change.
Location
Temperate forests throughout New Zealand.
Methods
We measured traits of foliar, stem, root and reproductive tissue for 64 species and calculated abundance-weighted mean trait values on 324 forest plots. Multiple linear regression was used to model the variation in each of the traits as functions of mean annual temperature (MAT), vapour pressure deficit (VPD), soil pH, soil total phosphorus (P) and their interactions.
Results
Soil–climate interactions explained significant variation in functional traits. For example, specific leaf area (SLA) was highest in high-P soil within a wet and warm climate; however, strong interactions indicate that SLA was lowest in wet and warm climates in low-P soil. Root tissue density was lowest in warm climates and high-P soil, but it was high in warm climates and low-P soil and in cold climates and high-P soil. According to model predictions, the largest potential responses of vegetation to warming may occur in fertile and wet environments.
Main conclusions
Pervasive soil–climate interactions demonstrate that interpreting simple bivariate relationships between traits and climate must be done with caution because the adaptive value of traits in a given climate depends on the fertility of the soil. Predictions of vegetation responses to climate change will improve significantly by incorporating local-scale soil properties into modelling frameworks. Rising global temperatures may shift community-level trait values in opposite directions depending on whether the soil is fertile or infertile.
Introduction
Over 110 years have passed since Schimper (1903) described how vegetation is driven by environmental factors, but only recently have ecologists been able to synthesize the multiple data sources that are needed to rigorously quantify how plant functional traits are related to multiple environmental gradients over large spatial scales (Violle et al., 2014; Maire et al., 2015). One of the core motivations for relating traits to environmental gradients is to quantify the adaptive value of traits in order to advance our understanding of how plant communities assemble over space and time (Lavorel & Garnier, 2002; Westoby & Wright, 2006; Laughlin & Messier, 2015). The influences of climate and soil fertility are particularly important because they are widely viewed as the primary factors driving plant survival and species distributions across biogeographical gradients (Prentice et al., 1992; Violle et al., 2014).
Mean annual temperature and moisture availability strongly relate to vegetation distributions because these factors affect rates of photosynthesis, transpiration, water and nutrient uptake, and growth (Jones, 1992; Meng et al., 2015). Temperature and precipitation vary in the extent to which they affect the distribution patterns of species and functional traits, and their interaction is likely to be significant (Moles et al., 2014). A conspicuous feature of trait–climate relationships is the large variation in trait values that occurs in any given climate, and much of this variation may be explained by local soil properties (Wright et al., 2004; Meng et al., 2015). Recent global studies have indicated that local variation in soil properties can exert stronger control over leaf traits than climate (Maire et al., 2015) and that some responses of leaf traits to soil fertility depend on climate (Ordoñez et al., 2009).
Predictable variation in plant community traits along individual gradients of either soil fertility (Richardson et al., 2004; Jager et al., 2015) or climate (Swenson & Weiser, 2010; Meng et al., 2015) has been demonstrated in many ecosystems, but climate and soil fertility may interact and cause non-additive effects on trait variability. For example, in annual plant communities, specific leaf area (SLA) increases with increasing nitrogen availability in moist soils, but it does not change in drier soils (Dwyer et al., 2015). At a global scale, the relationship between SLA and mean annual precipitation (MAP) changes depending on the fertility of the soil, probably because the constraints of soil fertility on leaf life span are more pronounced in dry climates (Ordoñez et al., 2009). Our predictions of the assembly and functioning of ecosystems in a rapidly changing world will depend on our understanding of how environmental factors interact to constrain ecosystem responses (Reich et al., 2006; Fridley et al., 2011; Peñuelas et al., 2013). The few studies that have addressed trait responses to interacting environmental gradients have provided compelling evidence that the adaptive value of a trait in a given climate depends on soil fertility, but our understanding of these interactions is currently limited to a few traits, namely leaf traits, maximum height and seed mass (Ordoñez et al., 2009; Dwyer et al., 2015; Maire et al., 2015). We know much less about the interacting effects of climate and soil on other plant properties, such as wood traits, root traits, bark thickness and flowering phenology.
Our overall aim was to determine how soil fertility and climate interact to simultaneously influence variation in functional traits representing all plant organs (foliar, stem, root and reproductive tissue). It was recently noted that local-scale data on soil properties measured in situ (rather than estimated from soil models) is needed in order to make more detailed assessments of soil–climate interactions (Maire et al., 2015). Our study has answered that call by synthesizing trait data, forest composition, climate and local-scale in situ soil data across temperate rain forests in New Zealand. We quantify how abundance-weighted mean traits were related to the interactions of mean annual temperature (MAT), vapour pressure deficit (VPD), soil pH and soil total phosphorus (P) to evaluate the importance of soil–climate interactions across biogeographical scales.
Methods
Study system
We tested trait–environment relationships across 324 plots spanning all major forest types in temperate rain forests throughout the main islands of New Zealand (see Fig. S1 in Appendix S1 in Supporting Information). Forest community composition was measured on 400-m2 plots (20 m × 20 m) at 324 LUCAS (Land Use and Carbon Analysis System) plot locations. Trees within each plot were identified to species, and diameters were measured on all individual stems ≥ 2.5 cm diameter at breast height (d.b.h.; 1.35 m).
The MAT and October VPD at each of the 324 plots were obtained from the Land Environments of New Zealand (LENZ) database (Leathwick, 2003) (Table 1). These climate variables were chosen because they have been shown to be the strongest predictors of species distributions throughout New Zealand and because they represent both temperature stress and moisture availability (Leathwick, 2001). VPD is calculated as [(100 – RH)e]/100, where RH is relative humidity and e is the saturation water vapour pressure (Leathwick, 2003). VPD is measured in October because during this month persistent westerly winds result in strong geographical variation in VPD across New Zealand. We investigated whether MAP or rainfall-to-potential evapotranspiration ratio (rainfall-to-PET ratio) would be a better predictor of trait variation, but models with VPD consistently yielded lower Akaike information criterion (AIC) scores than models with MAP or rainfall-to-PET ratio (Table S1).
| Variable | Abbreviation | Unit | Range |
|---|---|---|---|
| Environmental factors | |||
| Mean annual temperature | MAT | °C | 5.3–15.7 |
| October vapour pressure deficit | VPD | kPa | 0.01–0.54 |
| Soil pH | pH | –log[H+] | 3.04–6.25 |
| Soil total P | P | mg P kg−1 soil | 49–2360 |
| Functional traits | |||
| Specific leaf area | SLA | mm2 mg−1 | 1.7–25.6 |
| Leaf nitrogen concentration | LNC | %N | 0.62–3.32 |
| Leaf dry matter content | LDMC | g g−1 | 0.211–0.53 |
| Wood density | WD | mg mm−3 | 0.14–0.78 |
| Wood dry matter content | WDMC | g g−1 | 0.18–0.66 |
| Root tissue density | RTD | mg mm−3 | 0.08–0.31 |
| Specific root length | SRL | m g−1 | 6.69–130.52 |
| Maximum height | Htmax | m | 6–50 |
| Bark thickness | RBT | % | 0–6.19 |
| Flowering onset | Flronset | Julian date | 31–278 |
| Seed mass | SM | mg | 0.0008–1868 |
To assess soil fertility, nine mineral soil samples per plot were collected to a depth of 10 cm. Surface litter with botanically recognisable structures (the Oi and Oe layers) and the organic horizon (fermentation–humus layer) were scraped away before sampling. The organic horizon was removed to standardize all samples to the upper mineral horizon because not all sites had an organic horizon. The subsamples were pooled at the plot level prior to analysis. Soil pH was measured in solution with a 1:2.5 soil-to-distilled water ratio using a Radiometer PHM210 pH meter equipped with a Radiometer pHC2401-8 electrode. Soil total nitrogen (N) and organic carbon (C) were measured using a Leco CNS 2000 Analyser which utilises the Dumas dry combustion principle (Metson et al., 1979). Total P and organic P were determined using flow injection analysis on a Lachat QuikChem 8000 following ignition at 550 °C for 60 min and extraction with 0.5 m H2SO4 at a soil-to-extractant ratio of 1:200 for 16 h (Blakemore et al., 1987). We tested for total potassium (K) using atomic absorption spectroscopy with a Varian SpectrAA-220FS following leaching with molar ammonium acetate buffered to pH 7 (Blakemore et al., 1987).
From these measures, we selected soil pH and soil total P to represent gradients in soil fertility for two reasons. First, soil pH and available P were identified by Maire et al. (2015) as being the strongest predictors of leaf traits. We lack data on available P, but previous studies have shown strong positive correlations between total P and Olsen P in New Zealand soil (Maher & Thorrold, 1989; Parfitt et al., 2005). Second, pH and soil total P in our dataset loaded on separate principal components, and therefore are only weakly correlated soil properties (Fig. S2). Soil pH is regarded as a ‘master variable’ of soil chemistry (Binkley et al., 2011) because many mineral nutrients are unavailable in strongly acid soil; pH varied from strongly acidic (pH = 3) to only weakly acidic (pH = 6.3) (Table 1). Total P was also positively correlated with total N and K (Fig. S2), so variation in total P is also representative of variation in total pools of other macronutrients. Moreover, gradients in soil total P have been shown to have considerable influence on composition, structure and function of New Zealand rain forests (Richardson et al., 2004; Richardson et al., 2008).
Species and functional traits
We selected the 64 most frequent and abundant tree species in New Zealand forests across the national plot network based on the following criteria: the species must (1) attain a d.b.h. of at least 10 cm; (2) contribute at least 10% to the total basal area of at least one plot; and (3) occur on at least 1% of all the plots. We measured 11 functional traits on these species. In order to maximize trait sampling efficiency, sites were selected from across the broad latitudinal and altitudinal range in New Zealand (Fig. S1). This ensured that species were measured across a range of sites and that adequate intraspecific trait variation was captured. In total, trait data were collected from over 3400 individual trees.
We measured SLA, leaf dry matter content (LDMC) and leaf nitrogen concentration on all species. Mature, fully expanded, well-lit leaves from healthy canopies were sampled using a telescopic pruner or shotgun and were stored in sealed plastic bags. Leaf fresh mass was measured within 7 h of sample collection. Leaves were oven-dried at 60 °C for a minimum of 1 week before being reweighed to obtain oven-dry mass. LDMC was calculated as dry leaf mass divided by fresh leaf mass, and SLA was calculated as one-sided projected leaf area divided by dry mass. Leaf nitrogen concentration (% dry mass) was obtained using flow injection analysis on a Lachat QuikChem 8000 series at the Landcare Research Environmental Chemistry Laboratory, Palmerston North, New Zealand.
Wood traits were measured using different methods depending on the nature of the stems of each species. For angiosperms, conifers and the only palm tree (Rhopalostylis sapida), which constituted the vast majority of the tree species in this study, cores were extracted from the tree bole at breast height. For a few species where wood could not reliably be extracted with a borer (such as Dracophyllum spp.), a branch approximately 15 cm long was collected from a healthy lower branch. Data from another study was used to calculate stem traits of the tree ferns (Jager et al., 2015). Wood fresh mass measurements were made within 7 h of sample collection. Wood samples were oven dried at 100 °C for a minimum of 1 week before being reweighed. Wood dry matter content (WDMC) was calculated as the oven dry mass divided by wood fresh mass and wood density was calculated as wood dry mass divided by fresh volume. Bark thickness was measured using digital callipers on all extracted wood samples. The bark thickness of tree ferns and the palm tree was recorded as zero. Relative bark thickness, expressed as a percentage, was calculated as bark thickness divided by the d.b.h. (Lawes et al., 2014).
Root trait data were sourced from measurements on nursery-grown seedlings (Kramer-Walter et al., 2016). The root length of fine (< 2 mm) roots was measured using WinRhizo Pro software version 2012b (Regent Instruments Inc., Quebec City, Canada). The root sections were then dried at 60 °C for 48 h before obtaining dry mass. Specific root length (SRL) was calculated as root length divided by root dry mass, and root tissue density was calculated as root dry mass divided by fresh volume. We analyse both SRL and root tissue density here because they have been shown to be independent and may reflect different aspects of adaptation of plants to resource acquisition (Kramer-Walter et al., 2016).
Estimates of maximum plant height for each species were derived from regional floras (McGlone et al., 2010). Oven-dry seed mass was measured on each of the species from seeds stored in the Allan Herbarium (CHR; Lincoln, New Zealand). The average weight of a fern spore (0.0008 mg) was used as mean ‘seed mass’ for the six tree ferns in this study (Jager et al., 2015); however, these values did not significantly influence model results. The timing of onset of flowering for each species was extracted from the literature (Allan, 1961; Davies, 1961; Esler, 1969; Leathwick, 1984; Salmon, 1985; Best & Bellingham, 1991; Wilson, 1994; Dijkgraaf, 2002; Farjon, 2005). Dates were converted into Southern Hemisphere Julian dates, using 1 July as the first day of the Southern Hemisphere Julian calendar.
Data analysis
Our primary objective was to evaluate how soil–climate interactions explain functional trait variation at biogeographical scales (Violle et al., 2014), but plant traits, such as litter nutrient concentration and nutrient uptake rates can feed back to influence soil properties (Hobbie, 2015). It was recently shown that the environmental filtering of leaf traits by soil properties and the positive feedback effects of leaf litter on soil fertility can be equally strong processes in New Zealand forests (Laughlin et al., 2015). Here we focus exclusively on the process of environmental filtering and the nature of soil–climate interactions to improve predictions of vegetation properties across biogeographical scales.
Multiple linear regression was used to model each CWM trait as a function of the four environmental variables (soil pH, soil total P, MAT and VPD) and all their interactions using the lm function in R. All continuous predictor variables were centred and scaled to unit variance in order to properly estimate the interaction effects and significance of model terms (Aiken & West, 1991); however, the original variables were used to generate the final equations. The same following two-step procedure was used for each trait. First, the full four-way interaction model was fitted to the data, but if the four-way interaction term was not significant, then this term was removed from the model and refitted with all three-way interaction terms. If no three-way interaction terms were significant, then these were removed from the model and a two-way model was obtained. If no two-way interactions were significant, then a final model with only main effects was obtained. Second, to further evaluate the evidence for interactions, the second-order bias corrected AIC (AICc) of the final model that included interactions was compared with the AICc of a strictly additive main effects model to quantify the level of support for the interactions (Anderson, 2008). Differences in AICc (ΔAICc) > 7 between each model suggest that the model with the minimum AICc is more strongly supported. If the two models were equivalent, we report only the more parsimonious main effects model. To facilitate interpretation of the models, model-fitted surface plots were created using the ‘rsm’ library of R (Lenth, 2009), where the fitted predictions for each pair of environmental predictors is shown while holding the other environmental variables constant at their mean values.
Relationships among the four explanatory environmental gradients were evaluated using structural equation modelling (Grace, 2006) because climate has important long-term effects on soil development (Jenny, 1941). All bivariate correlation coefficients were < 0.5 among each pair of variables (Fig. S3A). The most parsimonious structural equation model indicated that soil pH and VPD tended to be higher in warmer environments, and soil total P tended to be higher in drier climates and in soils with higher pH (Fig. S3B). In other words, some of the warmest sites were also the driest, and soil total P and pH tended to be lowest in cold and wet climates, probably because mineral nutrients were leached from the soil in high-rainfall environments. Multicollinearity of these explanatory variables in multiple regression analyses was not problematic because all pairwise |r| < 0.5 and variance inflation factors (VIFs) for each term in the model were all less than three; terms with VIF > 10 are considered problematic in multiple regression models (Dormann et al., 2013).
We simulated scenarios with the final models to illustrate how the responses of vegetation to changes in temperature might depend on soil properties. We obtained regression model predictions of changes in average trait values due to a 3 °C rise in temperature from 10 to 13 °C under four scenarios: low soil fertility and low VPD; low soil fertility and high VPD; high soil fertility and low VPD; high soil fertility and high VPD. Low soil fertility was defined as soil pH = 4 and soil total P = 181 mg kg−1 (10th percentiles) and high soil fertility was defined as soil pH = 5.6 and soil total P = 897 mg kg−1 (90th percentiles). Low VPD was defined as 0.13 kPa (10th percentile) and high VPD was defined as 0.36 kPa (90th percentile).
Results
Bivariate scatterplots between each trait and each environmental predictor illustrate that soil pH, soil total P and MAT have stronger linear relationships with traits than VPD (Figs 1 & 2). However, these simple bivariate plots do not account for any potential interactions among factors. The multiple regression models that included interactions explained significant variation in community-weighted mean SLA (
= 0.40), LDMC (
= 0.47), leaf nitrogen concentration (
= 0.36), wood density (
= 0.16), WDMC (
= 0.29), root tissue density (
= 0.39), SRL (
= 0.41), maximum height (
= 0.18), bark thickness (
= 0.45), flowering onset (
= 0.22) and seed mass (
= 0.29) (Table 2). Four-way interactions were significant for three of the eleven models, at least one three-way interaction was significant in eight of the eleven models, and at least one two-way interaction between a soil and climate variable was significant in seven of the eleven models (Table 2). Based on ΔAICc, there was strong support for including interaction terms in all models except for SRL (Table 2). Given the ubiquity of interactions, the main effects cannot be interpreted directly and so we illustrate the results of the fitted models using model-fitted surface plots (Figs 3 & 4).
Bivariate scatterplots of five community-weighted mean (CWM) traits (specific leaf area, leaf dry matter content, wood dry matter content, root tissue density specific root length) and the four environmental variables (soil pH, soil total P, mean annual temperature, vapour pressure deficit). Correlation coefficients and line of best fit are shown for each of the bivariate relationships; dashed lines are not significant. ***P < 0.001; **P < 0.01; *P < 0.05.
Bivariate scatterplots of four community-weighted mean (CWM) traits (maximum height, bark thickness, flowering onset, seed mass) and the four environmental variables (soil pH, soil total P, mean annual temperature, vapour pressure deficit). Correlation coefficients and line of best fit are shown for each of the bivariate relationships; dashed lines are not significant. ***P < 0.001; **P < 0.01; *P < 0.05.
Model-fitted surface plots of community-weighted mean (CWM) traits (a–f, specific leaf area, mm2 mg−1; g–l, leaf dry matter content, mg mg−1; m–r, wood dry matter content, mg mg−1; s–x, root tissue density, mg mm−3; y–ad, specific root length, m g−1) along gradients of soil pH, soil total P, mean annual temperature and vapour pressure deficit.
Model-fitted surface plots of community-weighted mean (CWM) traits (a–f, maximum height, m; g–l, relative bark thickness, %; m–r, flowering onset, Julian date; s–x, seed mass, mg) along gradients of soil pH, soil total P, mean annual temperature and vapour pressure deficit.
| Term | SLA | LDMC | Leaf N | Wood density | WDMC | Root density | SRL | Maximum height | Bark thickness | Flowering onset | Seed mass |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Intercept | −23.22*** | 0.69*** | −4.79*** | −0.68*** | 0.22*** | 5.64*** | 5.60*** | 854.75*** | −9.30*** | 363.27*** | 298.01*** |
| pH | 15.58*** | −0.09*** | 2.32*** | 0.44* | 0.15n.s. | −1.24*** | −0.01n.s. | −191.27*** | 2.24** | −6.62* | −65.08* |
| Soil P (P) | 11.32*** | −0.12*** | 1.72*** | −0.02* | 0.10* | −0.79*** | −0.17*** | −132.14n.s. | −0.37*** | −24.07*** | −40.60n.s. |
| MAT | 4.90*** | −0.04*** | 0.99n.s. | 0.08* | −0.02*** | −0.74*** | −0.14*** | −113.63n.s. | 2.26*** | −55.76n.s. | −36.27*** |
| VPD | −344.80n.s. | 2.68*** | −40.01*** | 2.71n.s. | −1.07*** | −21.95n.s. | 1.93*** | −2808.00n.s. | 30.60*** | 564.24*** | −1081.51n.s. |
| pH × P | −4.04* | 0.04** | −0.54* | −0.02n.s. | −0.03** | 0.19** | 30.74n.s. | 0.15n.s. | −4.04n.s. | 8.78n.s. | |
| pH × MAT | −2.45n.s. | 0.02n.s. | −0.37** | −0.03*** | −0.0007n.s. | 0.17n.s. | 25.89n.s. | −0.37*** | 5.73*** | 7.91n.s. | |
| P × MAT | −1.65* | 0.02n.s. | −0.26* | 0.01n.s. | −0.0001n.s. | 0.11* | 18.46** | −0.11* | 6.07*** | 5.02n.s. | |
| pH × VPD | 48.11n.s. | −0.36n.s. | 5.87n.s. | −1.45n.s. | 0.07n.s. | 4.88*** | 593.46n.s. | −6.83n.s. | −68.38n.s. | 237.052n.s. | |
| P × VPD | 36.02** | −0.11n.s. | 4.79** | 0.13n.s. | 0.09n.s. | 3.30n.s. | 434.44n.s. | 2.31** | −25.12n.s. | 147.44n.s. | |
| MAT × VPD | 35.02n.s. | −0.35* | 3.45n.s. | −0.04n.s. | 0.04* | 2.85n.s. | 387.46* | −5.52n.s. | −22.93** | 132.23n.s. | |
| pH × P × MAT | 0.60* | −0.01n.s. | 0.08** | −0.001n.s. | −0.03n.s. | −4.24n.s. | 0.0001n.s. | −1.076n.s. | |||
| pH × P × VPD | −3.22n.s. | −0.02n.s. | −0.58n.s. | 0.11n.s. | −0.74n.s. | −93.20* | −0.32n.s. | −32.264n.s. | |||
| pH × MAT × VPD | −3.45*** | 0.05** | −0.31** | 0.10n.s. | −0.63*** | −82.37n.s. | 0.99*** | −28.81*** | |||
| P × MAT × VPD | −3.19* | 0.02n.s. | −0.32* | 0.10*** | −0.43n.s. | −60.88n.s. | 0.11n.s. | −18.28n.s. | |||
| pH × P × MAT × VPD | 0.09* | 13.096* | 3.97* | ||||||||
Model
 |
0.40 | 0.47 | 0.36 | 0.16 | 0.29 | 0.39 | 0.41 | 0.18 | 0.45 | 0.22 | 0.29 |
| F ratio | F14,309 = 16.5 | F14,309 = 21.8 | F14,309 = 13.7 | F14,309 = 5.3 | F10,311 = 13.8 | F15,308 = 14.7 | F4,309 = 55.0 | F15,308 = 5.7 | F14,309 = 19.9 | F10,261 = 8.5 | F15,306 = 9.8 |
| P value | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 | < 0.0001 |
| ΔAICc | 33.4 | 26.5 | 37.4 | 21.7 | 12.2 | 20.4 | 2.3a | 13.6 | 24.7 | 26.9 | 40.2 |
Main effects
 |
0.31 | 0.41 | 0.25 | 0.06 | 0.24 | 0.32 | 0.41 | 0.11 | 0.39 | 0.11 | 0.17 |
- The coefficients were obtained using unscaled data; however the significance of each coefficient was assessed where predictors were scaled to unit variance: ***P < 0.001; **P < 0.01; *P < 0.05; n.s., not significant. See Methods for a full explanation of the model selection procedure. ΔAICc > 7 indicate strong support for a model with interaction terms compared to an additive main effects-only model.
- aBased on the low ΔAICc for SRL, this model was reduced to a main effects-only model, despite evidence for a few significant interactions.
The leaf economic traits were strongly correlated among species (all |r|>0.69; Table S2). SLA and leaf nitrogen content were positively correlated, and these were both negatively correlated with LDMC. These three leaf traits responded similarly to the environmental gradients (Figs 3a–l & S4a–f). SLA was highest and LDMC was lowest in soil with higher pH and high total P, and in warm and wet climates (Fig. 3a–l), but interactions among soil and climate gradients indicate that these relationships were contingent on combinations of each factor (Table 2). For example, SLA was high in the warm and wet sites if soil pH was higher and was rich in total P, but SLA was low in the warm and wet sites if soil pH was low and poor in total P (Fig. 3a–l), indicating that soils regulate community-level leaf economic traits in a given climate.
Wood density was positively correlated with WDMC (r = 0.68) (Table S2), and responded similarly to environmental conditions, although less variation of wood density could be explained (Table 2). WDMC and wood density were both high in warm and dry sites but low in warm and wet sites (Figs 3q & S4j). WDMC was lowest in warm and wet environments, but rose quickly with increasing dryness or decreasing temperature (Fig. 3m–r). Wood density was lowest in warm sites with higher pH soil but was highest in warm sites with acidic soil (Fig. S4g–l).
Root tissue density was lowest in warm climates and high-P soil, but it was high in warm climates and low-P soil and in cold climates and high-P soil (Fig. 3s–x, Table 2). Root tissue density increased rapidly with decreases in P and pH (Fig. 3x). SRL was generally lowest in high-P soils and higher-pH soils and in warm and wet environments (Fig. 3y–ad). Given the lack of support for interaction terms based on a ΔAICc of 2.3, we report results of a model that includes only additive main effects. SRL was more related to soil P than to pH, where SRL was highest in low-P soil (Fig. 3ad).
Maximum height was generally highest in acidic soil and warm environments (Fig. 4a–f); however, there was a strong interaction between soil P and MAT (Table 2). In warm sites, soil P had a strong negative relationship with maximum height (Fig. 4a). The interaction between MAT and VPD indicated that plant height was low in both warm and wet sites and in cold and dry sites (Fig. 4e).
Bark thickness was generally lowest in high-pH and high-P soil, and in warm and wet climates (Fig. 4g–l); however, there were significant interactions between soil properties and climate (Table 2). Bark thickness could be high in high-P soil if the climate was cool and dry (Fig. 4g–l).
Flowering onset was earliest in low-pH and high-P soil, and in cool and dry climates (Fig. 4m–r); however, MAT interacted strongly with each of the other explanatory variables (Table 2). For example, flowering onset was early in both high P and cool sites and in low P and warm sites (Fig. 4m). Flowering onset tended to be earlier in wind-pollinated species (average = 108th Julian day), and latest in species with insect or vertebrate pollinators (average = 150th Julian day) (P = 0.08, Fig. S5), indicating that flowering phenology may be associated with both the environment and pollination syndrome.
Seed mass was largest in warm and wet sites with high soil P and higher pH (Fig. 4s–x, Table 2), but a significant four-way interaction indicates that the other factors interact in complex ways. For example, seed mass could be low in soil with high P if the climate was cool (Fig. 4s), or it could be low in warm climates if the pH of the soil was low (Fig. 4u).
The importance of these soil–climate interactions is illustrated by model predictions of potential responses of vegetation to a 3 °C rise in temperature from 10 to 13 °C in four scenarios of varying combinations of soil fertility and VPD (Fig. 5). SLA could respond positively in fertile and wet environments, but responses would be minimal in other environments (Fig. 5a). Wood density may increase slightly in infertile environments, but could decrease in fertile environments (Fig. 5b). Root tissue density could increase in dry sites with infertile soil, but is predicted to decrease in all other environments (Fig. 5c). Maximum height could increase on infertile soil in wet sites, but decrease on fertile soil in dry sites (Fig. 5d). Bark thickness will remain high on infertile soil, but could decline with warming on fertile soil (Fig. 5e). Flowering onset is predicted to be earlier with warming on infertile soil but is delayed with warming on fertile soil (Fig. 5f). Seed mass could increase with warming on fertile and wet sites, but changes would be limited in other conditions (Fig. 5g). Without exception, the largest absolute predicted changes in trait values to warming occurred in sites with fertile soil and a wet climate (Fig. 5).
Model predictions of shifts in community-level mean trait values after a 3 °C rise in mean annual temperature from 10 to 13 °C in four scenarios of different environmental conditions. Low (10th percentile) soil fertility was defined as soil pH = 4 and soil total P = 181 mg kg−1, and high (90th percentile) soil fertility was defined as soil pH = 5.6 and soil total P = 897 mg kg−1. Low VPD was defined as 0.13 kPa and high VPD was defined as 0.36 kPa.
Discussion
Our results show that the interactions between soil properties and climatic conditions are significantly related to community-level variation in functional traits from every plant organ across biogeographical gradients. These pervasive interactions demonstrate that interpreting simple bivariate relationships between individual traits and climate gradients must be done with caution because the observed variation of functional traits in a given climate depends on the fertility of the soil. The importance of these non-additive effects is receiving more attention (Ordoñez et al., 2009; Dwyer et al., 2015), and they have strong implications for many applications of trait-based ecology such as predicting the responses of species and vegetation to climate change across landscapes of varying soil properties (Reich et al., 2006). Our findings reinforce that predictions of the responses of vegetation to climate change will improve significantly by incorporating local-scale variation in soil properties in modelling frameworks. Our results indicate that the largest potential vegetation responses to warming may occur in fertile soil within wet climates, and that the responses of vegetation to warming may depend on the fertility of the soil (Fig. 5).
In sites with both high total P and higher pH, where stresses associated with nutrient limitation are minimized, variation in climate is related to variation in leaf traits. For example, in fertile soil LDMC varies from high in cool and dry environments to low in warm and moist environments. In fertile soil in cool and dry environments, ‘slow’ leaf traits (high LDMC, low SLA, low leaf N) enhance survival through a conservative strategy of building tough, long-lived leaves that are capable of photosynthesizing in harsh environments, and that have lower respiratory carbon and water loss than ‘fast’ leaves (Cunningham et al., 1999; Wright et al., 2005). It is only in fertile, warm and wet environments where ‘fast’ leaves (low LDMC, high SLA, high leaf N) are advantageous because the potential for rapid photosynthetic rates is not limited by any environmental factor (Grime, 1979; Wright et al., 2004). High LDMC is favoured in infertile soil across all climatic conditions because any reduction in cold and moisture-stress is insufficient to offset the detrimental effects of limitations in soil resources.
The multiple regression models determined whether soil and climate gradients could explain variation in plant traits, but we acknowledge that plant litter (i.e. leaves, wood, roots) can also feed back to influence soil properties and nutrient cycling (Hobbie, 2015). Plant–soil feedback is likely to make soils less fertile if the vegetation senesces leaves with high dry matter content and low nutrient content. Therefore, a limitation of these regression models is that they do not account for plant–soil feedback. Given that this study spans a large biogeographical gradient, the variation in soil fertility is likely to be influenced by geological substrates and topography (Werner & Homeier, 2015). However, vegetation history, including short-term successional dynamics as well as long-term soil development, has certainly influenced soil fertility (Jenny, 1941; Richardson et al., 2004). Previous work using non-recursive latent variable structural equation modelling has shown that the proficiency of leaves to resorb nutrients is engaged in positive feedback with soil fertility such that low-nitrogen litterfall reinforces low-fertility soil; however, wood density did not exhibit detectable feedback with soil but was strongly influenced by soil fertility (Laughlin et al., 2015). The analyses in this paper do not deny the importance of plant–litter feedbacks, but we acknowledge that the trait–soil relationships detected here are not solely driven by the soil but rather emerge from an interaction between plants and soil.
The model projections under warmer conditions (Fig. 5) must be considered in light of these known plant–soil feedbacks. If plants can influence the soil, then the constraints that soils place on vegetation under warmer conditions could be weaker than expected. If warming occurs faster than the time it takes for vegetation to influence soil properties, then plant feedback may be less important, but if these feedbacks are in sync with rates of warming, then our predictions of future responses must account for these plant-driven changes in soil properties. Understanding the differential rates of these temporal dynamics is an important area of global change research.
Wood density and WDMC are often described as having a negative relationship with soil fertility (Chave et al., 2009; Jager et al., 2015), but these relationships are contingent on climate. In fertile, warm and wet environments, species with low wood density are at an advantage because they have lower construction costs and can efficiently supply water and nutrients to a larger total leaf area to support rapid growth rates (Wright et al., 2010). However, in contrast to previous studies, high wood density can also be common on fertile soil if the climate is cool. Wood density and WDMC were high in dry environments that were either cool or warm because they confer resistance against both freezing- and drought-induced hydraulic cavitation, respectively (Sperry & Sullivan, 1992). We did not detect the widely observed positive relationship between wood density and temperature (Chave et al., 2009), for two reasons. First, temperate rain forests in New Zealand do not include the hottest environments at a global scale (Richardson et al., 2013); second, the value of wood density along temperature gradients depends on moisture availability. In fertile soil, low wood density was observed in the warmest environments as long as ample moisture was available. In infertile soil, high wood density is adaptive in any climate because it infers greater resistance to hydraulic cavitation (Sperry & Sullivan, 1992) and disturbances (Curran et al., 2008). Maintenance and repair costs are high in infertile soil, so being defended against such damage is advantageous in these environments.
Similar to leaves and wood, low root tissue density was most common in fertile, warm and wet sites. In these environments it is economical for species to produce low-density tissues that have low construction costs and provide a rapid return on investment over the short life span of a root (Freschet et al., 2013). High root tissue density is favoured in low-P and low-pH soil and in cool and dry climates where attaining sufficient resources is challenging. In these environments, species need to produce denser, longer-lasting tissue that is more resistant to belowground abiotic and biotic damage (Reich, 2014).
High SRL is an important adaptation for acquiring limited nutrients, especially P, because P becomes limiting once the diffusion shells surrounding root tissue are depleted (Holdaway et al., 2011; Laliberté et al., 2015). Indeed, SRL responded more strongly to gradients in soil total P than soil pH. However, SRL could also be high in high-P soil within cool and dry environments, which indicates that SRL may also be important for acquiring water when it is limited (Wasson et al., 2012).
Plant height is often thought to be greatest in resource-rich sites where limitations to growth and water transport are minimized (Givnish, 1995). However, in New Zealand, fertile sites are often occupied by species with a wide range of growth rates where competition for light is stronger. We speculate that this enables a greater range of niches related to shade tolerance and a greater range of maximum heights (Coomes et al., 2009), leading to lower community-weighted mean heights on fertile soil. Small-scale gap dynamics also occur frequently in these fertile sites making it favourable for species to allocate more resources to fast establishment rather than attaining tall height over the long term (Ogden et al., 1991). In contrast, infertile sites in New Zealand can sometimes be steep with thinner soil profiles. Less frequent, but larger scale, disturbances, such as landslides, are more common in these environments and slow-growing species that are able to tolerate these soils and attain height dominance are favoured (Wells et al., 2001), leading to higher community-weighted mean height. We acknowledge that this pattern may be unique to some New Zealand forests, because older soils around the world that have greater accumulations of organic matter can be both deeper and more acidic. Canopy height is reduced toward the tree line in cold subalpine environments, but calculations of community-level traits based on species-level averages do not reflect this widely observed trend (Fig. 2). Perhaps this is because maximum height, as measured by community-weighted means, may be less strongly related to physical environmental factors and instead more strongly influenced by local community dynamics and intraspecific variation in height.
Thick bark is typically viewed as a trait that confers resistance to fire, but recent work is showing that bark may confer other adaptive advantages (Jager et al., 2015; Richardson et al., 2015). Despite the lack of frequent fires in the evolutionary history of New Zealand trees (Lawes et al., 2014), thick bark has evolved in some trees which tend to be dominant in infertile soil regardless of climate. We speculate that thick bark in these environments may provide defence against other forms of disturbance such as abrasion (Richardson et al., 2015). In fertile soil, however, thin bark is common in warm and wet climates and thick bark is common in cool and dry climates, suggesting that thick bark may be advantageous for water storage (Rosell et al., 2014).
Flowering onset may be influenced by both the environment and by pollinator syndromes. Early onset of flowering in cool and dry environments may be a response to drought stress and the need to flower within a constrained growing season (Aronson et al., 1992). However, flowering phenology among species is also weakly associated with pollen-dispersal syndrome (Fig. S5). Early flowering species tend to rely on wind pollination, and wind speeds tend to be greatest in spring (Leathwick, 2003). Late-flowering species tend to be associated with bird, bat and insect pollinators (Fig. S5). These species probably have greater costs associated with investing in such pollinator mutualisms (nectar, fruits, large flowers) (Addicott, 1986), which may partly explain why many of these species occur in fertile, warm and wet environments.
Low seed mass was favoured in low-pH soil across all climates. However, in higher-pH soil, seed mass varied from low in cool and dry environments to high in warm and moist environments. Larger seeds may be favoured in fertile soil with warm and moist conditions where competition for light is highest because large seed size is positively associated with shade tolerance (Reich et al., 1998; Moles et al., 2007; Coomes et al., 2009). The association between large seeds and high temperatures in fertile sites may also be influenced by the increasing metabolic costs for respiration at higher temperatures (Murray et al., 2004). Low seed mass in infertile environments may reflect a strategy where production of high quantities of seed is favoured over the production of fewer large seeds (Leishman et al., 2000).
The most conducive environments for plant growth across New Zealand are warm and moderately wet climates with soils that contain high pools of mineral nutrients and higher pH (Wardle, 1991; Leathwick, 1995; Coomes et al., 2014). This combination of environmental conditions tends to be dominated by species with high SLA, low wood density, moderate root tissue density, moderate SRL, moderate maximum heights, thin bark, late flowering onset and moderate seed mass (Figs 3, 4 & S4). Many of these traits are associated with productive strategies of fast growth rates and low investment in defence (Reich, 2014). Perhaps the harshest environments for plant growth across New Zealand are infertile soils in cold and dry climates (Coomes et al., 2014). This combination of environmental conditions tends to be dominated by species with low SLA, high wood density, high root tissue density, high SRL, moderate maximum heights, thick bark, early flowering onset and low seed mass (Figs 3, 4 & S4). Many of these traits are typical of species that can tolerate stressful environments (Reich, 2014). However, we emphasize that these combinations of the four environmental variables are only two examples of environmental conditions, and that the fitness of multidimensional phenotypes varies continuously over the continuous changes in multidimensional environmental gradients (Figs 3 & 4). These complex and pervasive soil–climate interactions may explain some of the broad variation in traits that can be observed within a given climate (Wright et al., 2004; Meng et al., 2015) because the traits that confer fitness in a given climate depend on the fertility of the soil.
Incorporating soil–climate interactions into models of the distributions of species and vegetation will improve our understanding of community assembly processes and the prediction of vegetation distributions in a changing climate (Fridley et al., 2011). Climate and soil fertility have long been recognized as simultaneous drivers of vegetation distributions (Schimper, 1903; Prentice et al., 1992), but here, by analysing traits from every plant organ, we have shown that soil–climate interactions predict the distributions of multidimensional phenotypes. Predictions of vegetation distributions will improve by considering how multidimensional phenotypes respond to multidimensional environmental gradients (Laughlin & Messier, 2015). Soil and air water deficits have previously been shown to influence the distributions of tree species in New Zealand (Leathwick & Whitehead, 2001), but wide-ranging in situ data on soil fertility have not been available until now. This work highlights the importance of increasing the availability of high-resolution soil property data across biogeographical scales to improve predictions of ecological responses to climate change. The responses of species to changing temperature and precipitation regimes will be highly contingent on the local properties of the soil.
Acknowledgements
This research was supported by a grant (UOW1201) from the Royal Society of New Zealand Marsden Fund, a University of Waikato Research Scholarship and core funding for Crown Research Institutes from New Zealand's Ministry of Business, Innovation and Employment's Science and Innovation Group. We thank the New Zealand Department of Conservation's National Monitoring programme for collecting the soil samples at the national scale.
References
Biosketches
Angela H. Simpson is a plant ecologist interested in understanding vegetation distributions to inform sound ecosystem management.
Sarah J. Richardson is a plant ecologist interested in the biogeography of plant traits and the use of plant traits for interpreting vegetation change.
Daniel C. Laughlin is a plant ecologist who integrates principles of ecophysiology and plant community dynamics to improve our understanding and restoration of wild landscapes in a rapidly changing world.
