Grass–legume mixtures sustain strong yield advantage over monocultures under cool maritime growing conditions over a period of 5 years

Abstract

Background and Aims

Grassland-based livestock systems in cool maritime regions are commonly dominated by grass monocultures receiving relatively high levels of fertilizer. The current study investigated whether grass–legume mixtures can improve the productivity, resource efficiency and robustness of yield persistence of cultivated grassland under extreme growing conditions over a period of 5 years.

Methods

Monocultures and mixtures of two grasses (Phleum pratense and Festuca pratensis) and two legumes (Trifolium pratense and Trifolium repens), one of which was fast establishing and the other temporally persistent, were sown in a field trial. Relative abundance of the four species in the mixtures was systematically varied at sowing. The plots were maintained under three N levels (20, 70 and 220 kg N ha^–1 year^–1) and harvested twice a year for five consecutive years. Yields of individual species and interactions between all species present were modelled to estimate the species diversity effects.

Key Results

Significant positive diversity effects in all individual years and averaged across the 5 years were observed. Across years, the four-species equi-proportional mixture was 71 % (N20: 20 kg N ha^–1 year^–1) and 51 % (N70: 70 kg N ha^–1 year^–1) more productive than the average of monocultures, and the highest yielding mixture was 36 % (N20) and 39 % (N70) more productive than the highest yielding monoculture. Importantly, diversity effects were also evident at low relative abundances of either species group, grasses or legumes in the mixture. Mixtures suppressed weeds significantly better than monocultures consistently during the course of the experiment at all N levels.

Conclusions

The results show that even in the less productive agricultural systems in the cool maritime regions grass–legume mixtures can contribute substantially and persistently to a more sustainable agriculture. Positive grass–legume interactions suggest that symbiotic N₂ fixation is maintained even under these marginal conditions, provided that adapted species and cultivars are used.

INTRODUCTION

Agriculture is currently faced with the challenges of feeding an ever-increasing world population with fewer resources (Godfray et al., 2010), and the projected climate change will further accentuate the problem (IPCC, 2013; Thornton et al., 2014). Grassland-based livestock agriculture plays an important role as a provider of food across Europe, especially in the more marginal areas in the north, where arable production is limited (Helgadóttir et al., 2014). Such production systems are commonly dominated by grass monocultures receiving relatively high levels of fertilizer. It has been argued that basing these systems on grass–legume mixtures would improve their productivity and sustainability, with positive consequences for the environment (Lüscher et al., 2014). Here, we ask whether multispecies mixtures comprising selected species can improve productivity and robustness of yield persistence of cultivated grassland under extreme growing conditions over a period of 5 years.

There is a general consensus that biodiversity enhances ecosystem function (Hooper et al., 2005). In a pan-European Agrodiversity Experiment, it has been demonstrated that multispecies mixtures for intensively managed grasslands consistently give higher yields than would be expected from the yield of their component species in monoculture across a wide range of pedo-climatic conditions (Finn et al., 2013). Such diversity effects can be explained by improved acquisition of resources in time and space (niche complementarity), positive interspecific interactions (facilitation) and selection of the most productive species (sampling or selection effect) (Cardinale et al., 2007). Increased biomass production of mixtures can lead not only to overyielding (mixtures perform better than the average of component species grown in monocultures), but also to transgressive overyielding [the mixture yields more than its highest yielding component in monoculture (Trenbath, 1974)], where the latter can only be achieved through facilitation or improved resource acquisition (Cardinale et al., 2007). There is evidence that a priori selection of high-yielding species rather than species richness can explain increased yield and that strategic selection of the species of the mixtures is more important than the complexity of the mixtures (Sanderson et al., 2004; Lüscher et al., 2011; Storkey et al., 2015).

Species belonging to different functional groups are more likely to be complementary, as is the case with legumes and non-legumes. Even though two-species mixtures of legume and grass have been widely used in agriculture, there are indications that more diverse mixtures increase productivity (reviewed in Sanderson et al., 2004). Recent studies have shown that even a moderate increase in species number can result in increased ecosystem function such as biomass (Nyfeler et al., 2009; Finn et al., 2013) and total N yield (Nyfeler et al., 2011; Suter et al., 2015), when mixtures contained species belonging to a distinct functional group relating either to nitrogen (N) acquisition or to temporal development. Moreover, such mixtures grown at moderate N levels even showed transgressive overyielding and, at optimal species abundances, could yield more than excessively N-fertilized grass monocultures (Nyfeler et al., 2009). At moderately low levels of N fertilization, the diversity effects also seem robust to changes in the relative abundance of the mixture components, whereas at high N levels the diversity effects were reduced and even disappeared in the third year (Nyfeler et al., 2009). Besides negative effects of mineral N fertilizer on the activity of symbiotic N₂ fixation, this was primarily explained by the displacement of the legumes by the grasses, the latter being more competitive for N uptake and growth under high N conditions. Similarly, mixtures of species belonging to contrasting functional groups resisted the invasion of unsown species, probably as a result of better resource use (Frankow-Lindberg, 2012; Suter et al., 2017; Connolly et al., 2018).

The benefits of including fodder legumes in agricultural grasslands relate primarily to their ability to make use of atmospheric N₂ through symbiotic N₂ fixation, thus reducing their requirements for soil and fertilizer N (reviewed by Carlson and Huss-Danell, 2003; Lüscher et al., 2014). However, under marginal growing conditions in northern areas, symbiotic N₂ fixation is limited by low summer temperatures (Carlson and Huss-Danell, 2003; Rasmussen et al., 2013), and winter survival of the legume component is challenged by various stress factors such as low temperature, ice encasement and desiccation due to ground freezing (Bélanger et al., 2006; Helgadóttir et al., 2008). Indeed, in the previously mentioned European Agrodiversity Experiment, legume persistence was found to be negatively correlated with minimum site temperatures (Brophy et al., 2017). Probably because of these challenges with legume persistence, the performance of grass–legume mixtures is often unpredictable in the more marginal regions of Europe, both within and between years (Guckert and Hay, 2001), which may explain farmers’ reluctance to use such mixtures. When it comes to inter-annual yield stability, agronomically relevant studies with grassland mixtures rarely, if ever, extend beyond three harvest years. The latter would be particularly relevant in the more marginal areas of Northern Europe, where grass fields are commonly only renovated every 6–10 years.

The objective of this study is to quantify the effects of species richness and evenness of grass–legume swards on the yield and on weed invasion, and to assess yield persistence over a period of 5 years as affected by different N fertilization levels under cool maritime growing conditions. We adopted the approach used in the European Agrodiversity Experiment (Kirwan et al., 2014), which was further elaborated by Nyfeler et al. (2009). Two grass and two legume species were used in the current experiment, one of which was fast establishing and the other temporally persistent. These were sown at widely different species proportions and received three different N fertilizer application rates. We address the following questions. (1) Were grass–legume mixtures higher yielding and more resistant to weed invasion than can be expected from the performance of the individual species sown as monocultures? (2) Were diversity effects explained by N complementarity between grasses and legumes or by complementarity between fast establishing and temporally persistent species? (3) How did increased N fertilization affect the species’ relative abundance and observed diversity effects? (4) Were the diversity effects maintained over time, irrespective of N level?

MATERIALS AND METHODS

Experimental site

The experimental site was located at Korpa Experimental Station in South-West Iceland (64°09ʹ N, 21°45ʹ W, 29 m above sea level). The growing conditions at Korpa are characterized by relatively cool summers with long photoperiods. Meteorological data during the course of the experiment are summarized in Table 1. There was considerable variation in both temperature and precipitation, as well as incoming solar radiation, over the experimental period. Late spring and summer were relatively warm in year 2 but cold and wet with little sun in year 5.

Table 1.

Daily meteorological data for mean air temperature, soil temperature at 10 cm depth and solar radiation flux density, averaged over autumn (Au; 1 September–31 October), winter (Wi; 1 November–31 March), early spring (SpE; 1 April–30 April), late spring (SpL; 1 May–31 May) and summer (Su; 1 June–31 August), and total precipitation for the same periods during the course of the experiment, from summer 2008 (year 0) to autumn 2013 (year 5)

	Au	Wi	SpE	SpL	Su	Au	Wi	SpE	SpL	Su
	Mean air temperature (°C)					Mean soil temperature (°C)
Year 0					11.8					12.5
Year 1	5.8	0.7	5.0	7.5	11.6	6.2	0.2	2.5	7.3	12.6
Year 2	6.3	1.1	2.5	8.1	12.4	6.2	0.0	0.9	8.3	13.2
Year 3	7.9	0.2	4.3	6.9	10.8	7.7	-0.6	2.2	6.3	11.5
Year 4	6.9	1.2	4.3	6.3	11.8	6.3	0.9	3.7	6.8	13.1
Year 5	5.7	2.0	1.9	5.6	10.3	5.5	0.2	0.9	5.3	11.5
	Precipitation (mm)					Solar radiation flux density (W m–2)
Year 0					196					4534.5
Year 1	298	555	127	82	123	1414.3	559.1	2740.4	4904.5	4494.1
Year 2	193	351	42	32	147	1164.4	561.4	3049.5	4460.4	4116.7
Year 3	176	464	167	70	101	1357.2	377.7	2289.4	4462.6	4384.3
Year 4	353	735	75	25	146	1587.1	403.1	2952.2	5260.1	4713.1
Year 5	229	590	58	64	308	1597.9	558.0	3892.4	4269.5	3570.2

The soil at the site is typical Andosol (Gleyic Andosol) with strong andic soil properties, including low bulk density, strong phosphate retention, pH NaF of >11 and colloid chemistry dominated by allophane (10 %), ferrihydrate (8 %; giving 18 % total clay) and abundant metal humus complexes, which are partly responsible for the high organic content (8.3 % C) (methods by Blakemore et al., 1987) (Arnalds, 2015).

Experimental design and establishment

In June 2008, a field experiment was established on a total of 66 plots of 3 × 5 m. Four key forage species were chosen to represent different functional groups with regard to N acquisition and temporal development, namely two grass species, Phleum pratense ‘Snorri’ (G_Pp) and Festuca pratensis (syn. Schedonorus pratensis) ‘Norild’ (G_Fp), and two legume species, Trifolium pratense ‘Bjursele’ (L_Tp) and Trifolium repens ‘Norstar’ (L_Tr). The clovers fix N₂ from the atmosphere while the grasses take up N solely from the soil. Phleum pratense and T. pratense are considered to be fast establishing, while F. pratensis and T. repens are slow establishing but temporally more persistent. These functional types were selected, first to optimize complementarity of the species to maximize diversity effects in mixtures and, secondly to increase the stability of the mixtures over time. These species are also known to produce high-quality fodder in agricultural systems, and the selected cultivars are recommended for the site.

Species were sown following a simplex design (Cornell, 2002) which enables to evaluate a wide range of species relative abundance in the sward. Fifteen stands containing four monocultures and 11 mixtures were sown at two sowing densities (100 and 60 % of the recommended seeding rate). At full density, the seeding rates for monocultures of P. pratense, F. pratensis, T. pratense and T. repens were 20, 30, 14 and 12 kg ha^–1, respectively. Mixtures differed widely in sowing proportions: in four mixtures one species was dominant (70 % of the monoculture seeding rate of one species, 10 % of each of the other three), in six mixtures two species were co-dominant (40 % of each of two species, 10 % of the other two) and one mixture had an equal proportion of each species (25% of each) (Supplementary Data Table S1).

Three N fertilization levels were established to represent low, moderate and very high N levels for the management of agronomic grassland in Iceland. These were: 20 kg N ha^–1 year^–1 (N20), 70 kg N ha^–1 year^–1 (N70) and 220 kg N ha^–1 year^–1 (N220). The normal application of N to grass monocultures in Iceland is in the range of 120–150 kg ha^–1 year^–1. All plant stands were included in the N20 treatment (30 plots), whereas N70 and N220 consisted of the monocultures, the dominant stands and the equal stand (18 plots for each of these).

Maintenance and measurements

All plots received 40 kg P and 60 kg K ha^–1 year^–1 in early spring (around 10 May). Annual N fertilization was split into two applications for N70 and N220, where two-thirds were applied in early spring and one-third after the first harvest. For N20, all N fertilizer was applied in early spring. The plots were harvested twice a year for 5 years (2009–2013) using a hand-driven cutter bar mower (Agria 5300, Agria-Werke GmbH, Germany). Dry matter yield was determined at each harvest by drying sub-samples to a constant weight. Biomass proportions were determined by manual separation of plant samples from permanent sub-plots (50 × 50 cm) into the four sown species and the pooled unsown species portion for the first four harvest years. In the final year (2013), the yield was measured as before yet without botanical separation.

Data analyses

The yield data were analysed by multiple linear regression following principles established by Kirwan et al. (2009), Nyfeler et al. (2009) and Connolly et al. (2013). There, the community-level response is modelled as a linear combination of (1) identity effects of species as represented by their monoculture performance; (2) species net interactions, termed diversity effects (D: being positive, negative or neutral), which are defined as the difference between the actual mixture performance and that expected from the relative contribution of the constituent monocultures; and (3) any further variables, such as overall sowing density and N fertilization in our case.

Given this framework, a multistage procedure was applied to regress total (annual) biomass yield on the proportions of each sown species, overall sowing density (Dens) and level of N fertilization. Thus, a first, preliminary model was:

$$\begin{array}{l}\text{Yield}=\alpha \text{Dens}+{\beta }_{\text{1}}{P}_{\text{Pp}}+{\beta }_{\text{2}}{P}_{\text{Fp}}+{\beta }_{\text{3}}{P}_{\text{Tp}}+{\beta }_{\text{4}}{P}_{\text{Tr}}+{\gamma }_{\text{1}}{P}_{\text{Pp}}{\text{N}}_1+{\gamma }_{\text{2}}{P}_{\text{Fp}}{\text{N}}_1+{\gamma }_{\text{3}}{P}_{\text{Tp}}{\text{N}}_1+{\gamma }_{\text{4}}{P}_{\text{Tr}}{\text{N}}_1+{\gamma }_{\text{5}}{P}_{\text{Pp}}{\text{N}}_2\\ +{\gamma }_{\text{6}}{P}_{\text{Fp}}{\text{N}}_2+{\gamma }_{\text{7}}{P}_{\text{Tp}}{\text{N}}_2+{\gamma }_{\text{8}}{P}_{\text{Tr}}{\text{N}}_2+{\delta }_{\text{1}}{P}_{\text{Pp}}{P}_{\text{Fp}}+{\delta }_{\text{2}}{P}_{\text{Pp}}{P}_{\text{Tp}}+{\delta }_{\text{3}}{P}_{\text{Pp}}{P}_{\text{Tr}}+{\delta }_{\text{4}}{P}_{\text{Fp}}{P}_{\text{Tp}}+{\delta }_{\text{5}}{P}_{\text{Fp}}{P}_{\text{Tr}}+{\delta }_{\text{6}}{P}_{\text{Tp}}{P}_{\text{Tr}}+\epsilon \end{array}$$ (1)

where P denotes the species’ proportions in the stand, with Pp, P. pratense; Fp, F. pratensis; Tp, T. pratense; and Tr, T. repens. Equation (1) was fitted to yield at each individual year (year 1 to year 5) and to the average yield across the 5 years (years 1–5). The sowing proportions were used as predictors in year 1 and for the average yield across the 5 years, whereas the observed species proportions in annual biomass of the preceding year were used in the remaining years. Parameters β₁–β₄ estimate the effect of species’ proportional contributions on yield at N20 and, if P = 1, β coefficients estimate the yield of the species in monocultures. The variable N indicates different fertilizer levels (modelled as a factor), where N₁ = 1, if N fertilization was N70 and N₁ = 0 otherwise, and N₂ = 1 if N fertilization was N220 and N₂ = 0 otherwise. Parameters δ₁–δ₆ estimate the six possible pairwise interactions among the four species to assess diversity effects. The residual term ε is assumed to be normally distributed with constant variance σ².

Inspection of regression coefficients in eqn (1) revealed that overall sowing density was not significant in all years and across years (P = 0.134 in year 1, P > 0.2 in all other cases) and was omitted from all models. Moreover, to increase parsimony, a series of hierarchical models were constructed as described in detail in Nyfeler et al. (2009) to test whether the six pairwise species interactions could be grouped together to represent specifically interactions between grass and legume species (BGL) and within grass and legume species (WGL), i.e. the grass–grass and legume–legume interactions. Formally, it was finally tested whether:

$${\delta }_{\text{2}}{P}_{\text{Pp}}{P}_{\text{Tp}}+{\delta }_{\text{3}}{P}_{\text{Pp}}{P}_{\text{Tr}}+{\delta }_{\text{4}}{P}_{\text{Fp}}{P}_{\text{Tp}}+{\delta }_{\text{5}}{P}_{\text{Fp}}{P}_{\text{Tr}}={\delta }_{\text{I}}\left(P_{\text{Pp}}{P}_{\text{Tp}}+P_{\text{Pp}}{P}_{\text{Tr}}+P_{\text{Fp}}{P}_{\text{Tp}}+P_{\text{Fp}}{P}_{\text{Tr}}\right)={\delta }_{\text{I}}{D}_{\text{BGL}}$$

and

$${\delta }_{\text{1}}{P}_{\text{Pp}}{P}_{\text{Fp}}+{\delta }_{\text{6}}{P}_{\text{Tp}}{P}_{\text{Tr}}={\delta }_{\text{II}}\left(P_{\text{Pp}}{P}_{\text{Fp}}+P_{\text{Tp}}{P}_{\text{Tr}}\right)={\delta }_{\text{II}}{D}_{\text{WGL}}$$ (2)

Applying F-tests for comparison of models, it appeared that such averaging of δ parameters was generally parsimonious and justified, and pairwise species interactions were therefore grouped as:

$$D_{\text{BGL}}=P_{\text{Pp}}{P}_{\text{Tp}}+P_{\text{Pp}}{P}_{\text{Tr}}+P_{\text{Fp}}{P}_{\text{Tp}}+P_{\text{Fp}}{P}_{\text{Tr}}$$

$$D_{\text{WGL}}=P_{\text{Pp}}{P}_{\text{Fp}}+P_{\text{Tp}}{P}_{\text{Tr}}$$ (3)

Thus, an updated model, which also included interactions of D_BGL and D_WGL with the N fertilization treatment, was:

$$\begin{array}{l}\text{Yield }={\beta }_{\text{1}}{P}_{\text{Pp}}+{\beta }_{\text{2}}{P}_{\text{Fp}}+{\beta }_{\text{3}}{P}_{\text{Tp}}+{\beta }_{\text{4}}{P}_{\text{Tr}}+{\delta }_{\text{1}}{D}_{\text{BGL}}+{\delta }_{\text{2}}{D}_{\text{WGL}}+{\gamma }_{\text{1}}{P}_{\text{Pp}}{\text{N}}_1+{\gamma }_{\text{2}}{P}_{\text{Fp}}{\text{N}}_1+{\gamma }_{\text{3}}{P}_{\text{Tp}}{\text{N}}_1+{\gamma }_{\text{4}}{P}_{\text{Tr}}{\text{N}}_1\\ +{\gamma }_{\text{5}}{D}_{\text{BGL}}{\text{N}}_1+{\gamma }_{\text{6}}{D}_{\text{WGL}}{\text{N}}_1+{\gamma }_{\text{7}}{P}_{\text{Pp}}{\text{N}}_2+{\gamma }_{\text{8}}{P}_{\text{Fp}}{\text{N}}_2+{\gamma }_{\text{9}}{P}_{\text{Tp}}{\text{N}}_2+{\gamma }_{\text{1}0}{P}_{\text{Tr}}{\text{N}}_2+{\gamma }_{\text{11}}{D}_{\text{BGL}}{\text{N}}_2+{\gamma }_{\text{12}}{D}_{\text{WGL}}{\text{N}}_2+\epsilon \end{array}$$ (4)

where the parameter δ₁ estimates the average diversity effect by mixing grass and legume species (D_BGL) and δ₂ estimates the average diversity effect by mixing grass with grass and legume with legume species (D_WGL). It follows that δ₁D_BGL and δ₂D_WGL estimate overyielding in mixtures above the expectation from the average of monocultures. All other terms have meanings as explained for eqn (1).

Next, analysis of the raw data revealed that the legume proportion in mixture stands at N220 strongly decreased from the first year onward (Fig. 1C). Under this condition, the diversity effects D_BGL and D_WGL at N220 could not reasonably be estimated and it was decided to exclude all mixtures at N220 from regressions following eqn (4), but not the monocultures. Instead, it was tested by t-tests whether yield in mixtures at N220 would differ from yield in grass monocultures. Therefore, the final model applied to yield at each individual year became:

Fig. 1.

Proportion of the sown legume species from the total yield of mixtures of four species (A–C) from sowing to year 4 at three N fertilization levels, and proportions of the temporally persistent legume species (Trifolium repens) from the pooled yield of the two sown legumes (D, E) and of the temporally persistent grass species (Festuca pratensis) from the pooled yield of the two sown grasses (F–H). In (A–C), the means and the range from the minimum to the maximum of summed proportions of the two legumes are displayed when their totals in seed mixtures were 20, 50 and 80 %. In (D–H) Means and the minimum to maximum range are given for the temporally persistent species when it was in lower, equal or higher proportion in the seed mixture compared with the fast establishing species. At N70, (E), no data are available at lower proportion conditions due to lack of one legume, and at N220, proportions of the persistent legume species could not reasonably be computed due to low overall legume abundance (compare with C).

$$\begin{array}{l}\text{Yield }={\beta }_{\text{1}}{P}_{\text{Pp}}+{\beta }_{\text{2}}{P}_{\text{Fp}}+{\beta }_{\text{3}}{P}_{\text{Tp}}+{\beta }_{\text{4}}{P}_{\text{Tr}}+{\delta }_{\text{1}}{D}_{\text{BGL}}+{\delta }_{\text{2}}{D}_{\text{WGL}}+{\gamma }_{\text{1}}{P}_{\text{Pp}}{\text{N}}_1+{\gamma }_{\text{2}}{P}_{\text{Fp}}{\text{N}}_1+{\gamma }_{\text{3}}{P}_{\text{Tp}}{\text{N}}_1+{\gamma }_{\text{4}}{P}_{\text{Tr}}{\text{N}}_1\\ +{\gamma }_{\text{5}}{D}_{\text{BGL}}{\text{N}}_1+{\gamma }_{\text{6}}{D}_{\text{WGL}}{\text{N}}_1+{\gamma }_{\text{7}}{P}_{\text{Pp}}{\text{N}}_2+{\gamma }_{\text{8}}{P}_{\text{Fp}}{\text{N}}_2+{\gamma }_{\text{9}}{P}_{\text{Tp}}{\text{N}}_2+{\gamma }_{\text{1}0}{P}_{\text{Tr}}{\text{N}}_2+\epsilon \end{array}$$ (5)

Here, D_BGL, D_WGL, and interactions of D_BGL and D_WGL with the N fertilization treatment were only included in the model if significant at P < 0.05.

In eqns (4) and (5), the diversity effects D_BGL and D_WGL are linearly related to the pairwise products of species proportions in a stand (Kirwan et al., 2009). This, however, must not necessarily be the case, as a species can have a profound effect on biomass yield even at comparably low relative abundance in a mixture. To test whether D would be non-linearly related to the pairwise species interactions, we used a generalized diversity–interaction model following Connolly et al. (2013) and estimated the pairwise interactions raised to the power of an additional coefficient θ (see Connolly et al., 2013 for an in-depth discussion on the meaning of the theta parameter). Only the data for averaged yields across the 5 years were used for this regression to increase reliability of the outcome. Thus, for years 1–5, D_BGL and D_WGL in eqn (5) were extended to:

$$D_{\text{BGL}\_\text{Theta}}={\left(P_{\text{Pp}}{P}_{\text{Tp}}\right)}^{\theta }+{\left(P_{\text{Pp}}{P}_{\text{Tr}}\right)}^{\theta }+{\left(P_{\text{Fp}}{P}_{\text{Tp}}\right)}^{\theta }+{\left(P_{\text{Fp}}{P}_{\text{Tr}}\right)}^{\theta }$$

$$D_{\text{WGL}\_\text{Theta}}={\left(P_{\text{Pp}}{P}_{\text{Fp}}\right)}^{\theta }+{\left(P_{\text{Tp}}{P}_{\text{Tr}}\right)}^{\theta }$$ (6)

with the regression model otherwise being equivalent to eqn (5). The size of the parameter θ was estimated by profile likelihood (Pawitan, 2001). More details of the regression models are given in Supplementary Data Appendix A.

Transgressive overyielding (mixture yield being higher than the highest yielding monoculture) was evaluated following Nyfeler et al. (2009). Based on eqn (5), it was tested for a range of predicted mixture yields whether these would significantly outperform the best yielding monoculture. See Supplementary Data Appendix A for details on computation, and Nyfeler et al. (2009) for an extended discussion on the applied method.

Biomass of unsown species was analysed with similar regression principles as explained for eqns (1)–(4) but using a generalized linear model to account for the greatly differing biomass of unsown species in stands. Here, we were mainly interested to know whether mixtures on average would suppress unsown species more than monocultures. In a similar way, we analysed proportions of unsown species (biomass of unsown species divided by the total biomass yield of a stand) using beta regression (Cribari-Neto and Zeileis, 2010). The specific distributions, link functions and regression equations are given in Supplementary Data Appendix A. All analyses were performed using the statistical software R (R Core Team, 2017).

RESULTS

High and robust overyielding and transgressive overyielding by mixing grass and legume species

Mixing grass and legume species (BGL effects) resulted in significant positive diversity effects in all individual years and averaged across the 5 years (Table 2). Positive diversity effects also appeared through the grass–grass and legume–legume interactions (WGL effects) but these were significant only in years 3 and 4 and were approximately half the size of the BGL effects. The relative magnitude of total overyielding (BGL plus WGL) was largest in year 3 (119 % at N20, 79 % at N70), yet the relative degree of overyielding was generally maintained over the 5 years. Averaged across the 5 years, overyielding by mixing grasses and legumes (BGL) was 71 and 51 % at N20 and N70, respectively, but was not significant by mixing grass with grass and legume with legume species (WGL) (Table 2). The theta parameter, indicating a non-linear diversity effect on yield, was estimated to be 0.305 (χ² = 7.03, P = 0.008, across the 5 years). Thus, a profound diversity effect was also evident due to mixing grasses and legumes at a low proportion of either species group in the mixture (Fig. 2C).

Table 2.

Predicted total yield (including unsown species) of monocultures, predicted yield and overyielding of the equal stand mixture, and predicted yield, transgressive overyielding and legume proportion of the mixture with highest yield, for the three N fertilization levels at five harvest years and averaged across the 5 years (1–5)

Year		Monocultures					Equal stand mixture‡					Mixture with highest yield‡
								Overyielding§				Yield	Transgressive overyielding		Legume proportion¶
		Pp	Fp	Tp	Tr	Average	Yield	BGL effect		WGL effect
		(t ha–1 year–1)					(t ha–1 year–1)	(t ha–1 year–1)	(%)	(t ha–1 year–1)	(%)	(t ha–1 year–1)	(t ha–1 year–1)	(%)	(%)
1	N20	2.68	4.05	5.25	4.20	4.05	6.29	2.25***	56	•	•	6.94	1.68*	32	57
	N70	4.85	5.54	5.12	4.68	5.05	7.30	2.25***	45	•	•	7.58	2.04**	37	48
	N220	7.18	8.18	5.75	5.17	6.57	–	–	–	–	–	–	–	–	–
	s.e.	0.511	0.511	0.511	0.511	0.252	0.236	0.323				0.343	0.449
2	N20	4.23	4.91	4.38	2.30	3.95	6.18	2.22***	56	•	•	6.87	1.97**	40	47
	N70	7.01	6.39	5.08	3.42	5.47	7.70	2.22***	41	•	•	8.37	1.36†	19	39
	N220	9.59	11.07	5.96	5.46	8.02	–	–	–	–	–	–	–	–	–
	s.e.	0.486	0.546	0.526	0.528	0.251	0.235	0.292				0.345	0.419
3	N20	1.69	1.67	4.31	2.68	2.59	5.68	2.03***	78	1.07*	41	5.98	1.67†	39	74
	N70	3.90	3.57	5.20	3.06	3.93	7.03	2.03***	52	1.07*	27	7.18	1.98*	38	55
	N220	6.87	8.06	5.43	4.70	6.27	–	–	–	–	–	–	–	–	–
	s.e.	0.443	0.505	0.495	0.527	0.237	0.293	0.427		0.453		0.368	0.469
4	N20	2.68	2.85	6.31	4.05	3.97	6.27	1.53***	38	0.77*	19	6.96	0.65n.s.	10	90
	N70	4.74	4.57	6.82	4.35	5.12	7.42	1.53***	30	0.77*	15	7.74	0.92n.s.	13	65
	N220	8.13	8.78	5.72	4.84	6.87	–	–	–	–	–	–	–	–	–
	s.e.	0.417	0.443	0.397	0.443	0.201	0.195	0.385		0.369		0.395	0.390
5	N20	0.68	1.43	1.91	1.77	1.45	2.27	0.83*	57	•	•	2.51	0.60n.s.	31	57
	N70	3.51	2.74	3.10	3.13	3.12	3.95	0.83*	26	•	•	4.16	0.65n.s.	19	44
	N220	4.97	5.48	5.10	4.06	4.90	–	–	–	–	–	–	–	–	–
	s.e.	0.583	0.581	0.490	0.580	0.262	0.252	0.310				0.345	0.477
1–5	N20	2.52	3.21	4.13	2.90	3.19	5.44	2.25***	71	•	•	5.63	1.50***	36	67
	N70	4.45	4.59	4.86	3.80	4.43	6.68	2.25***	51	•	•	6.75	1.89***	39	51
	N220	7.35	8.31	5.59	4.85	6.53	–	–	–	–	–	–	–	–	–
	s.e.	0.312	0.312	0.312	0.312	0.161	0.137	0.191				0.158	0.263

All values including standard errors (s.e.) are based on multiple linear regression [eqn (5)], with yield of the highest yielding mixture being predicted from the model. The maximal monoculture yield at each N level is in bold. Standard errors are computed from the weighted average of variances at each N fertilization level.

Pp, P. pratense; Fp, F. pratensis; Tp, T. pratense; Tr, T. repens.

^‡Mixture predictions at N220 could not reasonably be computed due to a strong shift of legume proportion (see Fig. 1C). Respective table cells are marked with a dash (–).

^§Overyielding is calculated as the difference between the yield of the equal stand mixture (i.e. mixture with an equal proportion of the four species) and the average yield of the four monocultures (Average). Overyielding is split into the effect of mixing grasses with legumes (between grass legume, BGL) and the pooled effect of grass–grass and legume–legume mixing (within grass legume, WGL); entire overyielding corresponds to the sum of these two components. Predictions and the level of significance are based on multiple linear regression [eqn (5); Supplementary Data Table S2].

^¶Legume proportion of the regression predictor: year 1 and across the 5 years (1–5): sown legume proportion; individual years 2–5: observed legume proportion of previous year.

***P ≤ 0.001; **P ≤ 0.01; *P ≤ 0.05; ^†P ≤ 0.1; n.s., not significant (inference based on t-test); •, WGL effect not significant and omitted from the model.

Fig. 2.

Predicted total yield (including unsown species) of mixtures (±1 s.e., light grey shaded) as a function of legume proportion for two N fertilization levels (N20 and N70) in the first year (A), the fourth year (B) and averaged across 5 years (C), including the predicted yield of the highest yielding mixture. Lines and s.e.s are based on regression analysis [Table 2; eqn (5)] and are displayed for mixtures with equal proportions of the two grass species and the two legume species, meaning that the left and right endpoints of the lines are binary mixtures and the prediction at 50 % legume proportion represents the four-species equi-proportional mixture (see the Materials and Methods for details of the design). Predictions for mixtures at N220 could not reasonably be computed due to a strong shift of legume proportion (Fig. 1C). Horizontal lines at the bottom of the panels indicate the range of significant transgressive overyielding of the predicted lines [P ≤ 0.1 in (A), P ≤ 0.05 in (C)]. Predicted monoculture yield (±1 s.e.) is indicated by symbols on either side of the mixture line; non-visible s.e.s are due to small values. Note that predictions in (C) are based on a regression that included the theta parameter [compare eqns (5) and (6)]. N20 = 20 kg N ha^–1 year^–1; N70 = 70 kg N ha^–1 year^–1, N220 = 220 kg N ha^–1 year^–1.

Equally important, we observed transgressive overyielding in the first 3 years between 32 and 40 % at N20, and between 19 and 38 % at N70 (Table 2). Averaged across the 5 years, transgressive overyielding of the highest yielding mixture was 36 and 39 % at N20 and N70, respectively, and was significant over a large range of legume proportions in mixtures, ranging from 43 to 90 % at N20, and from 15 to 93 % at N70 (Fig. 2C). Because transgressive overyielding across the 5 years was due to a strong BGL effect including the theta parameter, this indicates a robust benefit of mixing grass with legume species despite considerable compositional changes in mixtures.

At N220, yields of stands sown as mixtures were never significantly different from the yields of grass monocultures, either in individual years or across the 5 years (Supplementary Data Fig. S2). This was explained by the strong compositional shift of mixtures at N220, leading to compositions that mainly contained grasses (Fig. 1C).

At N20, the maximum yielding monoculture was the legume T. pratense in all years, except in year 2. In comparison, at N220, the maximum yielding monoculture was the grass F. pratensis, except in year 5 (Table 2). Thus, there was a larger N fertilization effect on yields of the grass species than on the legume species (Table 2; Fig. 2; Supplementary Data Table S2).

Mixtures suppressed weeds

We found generally less biomass of unsown species on average in mixtures than in monocultures at all N levels in the first 4 years and averaged across the 4 years (no botanical analyses were available for the fifth year) (Table 3), i.e. mixtures were more resistant to weed invasion compared with monocultures. The suppressive effect of mixtures was more pronounced when evaluated on the proportional scale, for each of the 4 years and averaged across years (P < 0.001 in all cases; Supplementary Data Table S3). This was mainly due to less biomass of unsown species but also to higher total biomass yield in mixtures than in averaged monocultures (compare Tables 2 and 3).

Table 3.

Predicted biomass of unsown species (kg ha^–1 year^–1) of monocultures, of the monocultures’ average and of the equal stand mixture (as a reference mixture) for the three N fertilization levels at four harvest years and averaged across the 4 years (1–4)

		Year 1	Year 2	Year 3	Year 4	Year 1–4
N20	P. pratense	403 (116.5, 163.8)*	337 (139.1, 237.1)	197 (50.1, 67.2)	870 (234.4, 320.9)	513 (99.5, 123.4)
	F. pratensis	1447 (417.9, 587.6)	1747 (728.7, 1250.3)	590 (153.0, 206.5)	809 (215.7, 294.1)	1055 (204.5, 253.6)
	T. pratense	1913 (552.5, 776.8)	1074 (425.1, 703.4)	1736 (416.5, 548.0)	2077 (474.4, 614.8)	1810 (350.9, 435.3)
	T. repens	3533 (1020.6, 1435.1)	3219 (1292.6, 2159.9)	1254 (332.7, 453.0)	3706 (985.0, 1341.4)	2532 (490.8, 608.8)
	Monoculture average	1409 (248.7, 302.0)	1194 (290.6, 384.0)	709 (101.6, 118.6)	1526 (221.1, 258.5)	1255 (145.0, 163.9)
	Equal stand mixture	469 (67.0, 78.2)	106 (21.8, 27.4)	545 (87.6, 104.4)	690 (88.7, 101.8)	399 (37.1, 40.9)
	t-value	3.95	6.24	1.03	3.60	6.51
	P-value	< 0.001	< 0.001	0.308	< 0.001	< 0.001
N70	P. pratense	617 (185.3, 264.7)	187 (66.3, 102.8)	569 (125.7, 161.3)	1070 (271.2, 363.3)	595 (120.0, 150.4)
	F. pratensis	2198 (659.6, 942.3)	1024 (429.1, 738.2)	717 (181.4, 242.7)	1174 (312.5, 426.0)	1316 (265.7, 332.9)
	T. pratense	2088 (626.5, 895.0)	1510 (639.9, 1110.4)	2737 (724.8, 985.9)	2882 (768.3, 1047.6)	2082 (420.3, 526.6)
	T. repens	3242 (972.8, 1389.7)	3207 (1337.6, 2294.5)	1985 (529.1, 721.3)	3164 (854.9, 1171.3)	2733 (551.7, 691.2)
	Monoculture average	1741 (309.8, 376.9)	981 (232.2, 304.2)	1220 (171.4, 199.4)	1840 (269.3, 315.5)	1453 (169.2, 191.5)
	Equal stand mixture	975 (220.6, 285.1)	345 (114.3, 170.9)	513 (109.5, 139.3)	1217 (211.1, 255.5)	573 (85.8, 100.9)
	t-value	1.60	2.04	2.81	1.57	4.07
	P-value	0.115	0.046	0.007	0.122	<0.001
N220	P. pratense	1748 (524.3, 749.1)	363 (123.7, 187.6)	1214 (239.5, 298.4)	1572 (381.4, 503.7)	1486 (299.9, 375.8)
	F. pratensis	2901 (870.3, 1243.4)	2469 (1039.4, 1794.8)	2581 (668.1, 901.4)	2682 (723.3, 990.4)	2604 (525.6, 658.6)
	T. pratense	4405 (1321.6, 1888.0)	6008 (2553.3, 4440.1)	4918 (1313.6, 1792.3)	4471 (1218.9, 1675.7)	4254 (858.8, 1076.0)
	T. repens	3976 (1193.0, 1704.4)	5026 (2160.6, 3789.8)	4274 (1145.5, 1564.8)	3804 (1038.7, 1428.9)	4143 (836.4, 1048.0)
	Monoculture average	3070 (546.2, 664.5)	2281 (535.2, 699.3)	2849 (391.6, 454.0)	2910 (420.6, 491.6)	2873 (334.8, 378.9)
	Equal stand mixture	972 (219.8, 284.)	84 (34.8, 59.5)	501 (141.1, 196.4)	654 (178.7, 245.9)	417 (62.4, 73.4)
	t-value	3.18	5.23	4.51	3.90	8.44
	P-value	0.003	<0.001	<0.001	<0.001	<0.001

All values including standard errors (in parentheses) are back-transformed from generalized linear regression [eqn (S1); Supplementary Data Appendix A]. The inference (t- and P-values) refers to the difference in predicted unsown species’ biomass between the monocultures’ average and the equal stand mixture [test of D_AVE, eqn (S1), Supplementary Data Appendix A], and indicates whether mixtures on average suppressed unsown species more than monocultures. For the proportional values of unsown species, see Supplementary Data Table S3.

N20 = 20 kg N ha^–1 year^–1, N70 = 70 kg N ha^–1 year^–1, N220 = 220 kg N ha^–1 year^–1.

*Standard errors (lower s.e., upper s.e.) are asymmetric due to back-transformation.

Legume proportion in mixtures sustained over time at low to moderate N fertilization levels

The summed proportions of both legumes in mixtures in year 4 were on average 70 and 39 % at N20 and N70, respectively (Fig. 1A, B), indicating that these legume species could be sustained at moderate N fertilization levels in these marginal areas of the north. When assessing the proportion of the fast establishing vs. the temporally persistent legume species, contrary to expectation, the legume T. repens did not increase in relative abundance over time (Fig. 1D, E). In comparison, proportions of the fast establishing legume T. pratense were sustained at N20 and N70, even when sown in sub-dominant proportions, thus behaving rather as a temporally persistent species (Fig. 1D, E: T. pratense proportion = 1 minus the proportion of the persistent legume species). Regarding the two grasses, as anticipated, P. pratense established rapidly but generally decreased in relative abundance over time, while proportions of the more persistent F. pratensis increased (Fig. 1F–H).

At N220, the proportions of the legume species after 4 years were 3 % each while the two grass species dominated all mixture types irrespective of original sowing proportions: on average, the proportions of P. pratense and F. pratensis were 47 % each.

DISCUSSION

Strong diversity effects in grass–legume mixtures sustained over 5 years and a wide compositional range at moderate N levels

Our results demonstrate that the cultivation of grass–legume mixtures in a cool, maritime climate resulted in sustained yield advantages at moderate N levels compared with the cultivation of the component species in monoculture. Across the 5 years, overyielding of the equal stand mixture was 71 and 51 % (compared with averaged monocultures) receiving 20 and 70 kg N ha^–1 year^–1, respectively, and for transgressive overyielding the yield ratios of the highest mixture/highest monoculture were 1.36 and 1.39 (Table 2). These are much higher values than were observed on average across sites and years in the European Agrodiversity Experiment presented by Finn et al. (2013), who reported a 32 % yield advantage of mixtures over averaged monocultures and transgressive overyielding of 7 % at moderate N levels. Likewise, our mixture benefits were considerably higher than those observed at sites experiencing fairly extreme climates in northern Europe and Canada (Sturludóttir et al., 2014). In the present experiment, the same species were used as in these two studies, except that F. pratensis was sown as the temporally more persistent grass species instead of Poa pratensis. Festuca pratensis is intrinsically higher yielding than Poa pratensis and with longer leaves. It should therefore be able to intercept more light in a dense canopy than Poa pratensis and, thus, be capable of higher resource acquisition in a mixture with P. pratense.

Diversity effects were primarily driven by the grass–legume interaction and were significant over a period of 5 years, even though the relative magnitude fluctuated between years. Assuming that the diversity effects were linearly related to the grass–legume interactions (Kirwan et al., 2009), averaged across the 5 years, transgressive overyielding was significant in mixtures with legume proportions between 42 and 73 % (N20) and between 25 and 80 % (N70) (Supplementary Data Fig. S3). However, testing the generalized diversity–interaction model (Connolly et al., 2013), it was revealed that the diversity effect was non-linearly related to the grass–legume interactions (Fig. 2C), implying that diversity effects must have also been evident at lower proportions of either species group in the mixture (e.g. at N70 transgressive overyielding was significant at around 10 % of either grass or legume in the mixture). Concluding from the χ² value (7.03), these diversity effects at low species proportions must be substantial (compare Fig. 2C and Supplementary Data Fig. S2).

Considerable variation in total yield was observed between years. Thus, in year 3 total yield was clearly lower than in both the years before and after, and in year 5 total yield more or less collapsed (Table 2). In both years 3 and 5, soil temperatures in spring were relatively low and the summer was particularly cold and wet in year 5 with low levels of solar radiation flux density (Table 1). However, the relative diversity effects, particularly at N20, were maintained and, indeed, the largest relative diversity effects were observed in year 3, implying that positive species interactions were not reduced in spite of increased stress.

Diversity effects are mainly driven by positive interactions between legumes and grasses

The predominant effect explaining the observed diversity response (BGL effect) is most probably the symbiotic N₂ fixation in legumes. In grass–legume mixtures of productive forage grassland, Nyfeler et al. (2011) have shown that the percentage N derived from symbiosis in T. pratense and T. repens was maximal when the legume proportion was between 20 and 40 %, and that the companion grasses have increased N contents as compared with grass monocultures. Taken together, these findings imply increased total N yields in well-balanced grass–legume mixtures. In agreement with this, it was recently demonstrated in a multi-site study that these mixtures attain their maximum total N yield when the legume component makes up around one-third of the mixture, with few to no further benefit from a higher legume proportion (Suter et al., 2015). This results from strong stimulation of symbiotic N₂ fixation driven by N demand of the companion grass (Høgh-Jensen and Schjoerring, 1997), being further accentuated when grasses are dominant in the mixture (Nyfeler et al., 2011). In our study, the significant non-linear diversity effect on average yield across the 5 years at moderate N levels is in agreement with these findings, in that fairly low proportions of legumes (i.e. with grasses dominant) have contributed disproportionately and substantially to the diversity effect.

The relatively high diversity effects observed in the current study under moderate N levels are of interest in light of the prevailing climatic conditions at Nordic sites, which are generally considered to be unfavourable for perennial legumes (Helgadóttir et al., 2008; Brophy et al., 2017). From nutrient solution experiments with white clover ecotypes, it has been concluded that, although temperatures in the range of 5–10 °C significantly reduced symbiotic N₂ fixation, reduced biomass accumulation was due to the indirect result of little shoot growth and, hence, low demand for N (Kessler et al., 1990; Svenning and Macduff, 1996; Hartwig, 1998). Also, in a recent field study at the current experimental site, it was demonstrated that white clover in an unfertilized grass mixture derived up to 99 % of N from symbiotic N₂ fixation (%Ndfa) (Rasmussen et al., 2013). This is a considerably higher fraction than is generally found elsewhere in the northern temperate/boreal areas (average around 80 %) (Carlson and Huss-Danell, 2003), but consistent with native and locally adapted legume species in the Alps where symbiotic N₂ fixation did not decline with altitude (Jacot et al., 2000a, b). It seems therefore that, with adapted cultivars (Svenning and Macduff, 1996) and as long as a minimum proportion of legumes is maintained in the sward, their potential to fix nitrogen is not hampered per se by the relatively low temperatures experienced during the growing season.

Significant grass–grass and legume–legume interactions were also observed for 2 years (Table 2). Grass–grass interactions, without the presence of legumes, were predicted to give around 1 t dry matter (DM) ha^–1 in year 4 at both N20 and N70 (Fig. 2B). Indeed, in a study with a four-species grass–legume mixture in Norway containing Lolium perenne and Festuca arundinacea, the most persistent positive species interactions were between the two grass species (Ergon et al., 2016). The positive grass interactions observed in our experiment could be additive effects due to niche complementarity regarding seasonal asynchrony in shoot growth (Husse et al., 2016) and/or rooting depth (Husse et al., 2017). Cultivar trials across the West Nordic countries and Sweden have clearly demonstrated that adapted cultivars of F. pratensis recover more rapidly after defoliation than those of P. pratense (Thorvaldson et al., 2015). The role of different rooting depths is less conclusive. However, F. pratensis is generally considered to be more drought tolerant than P. pratense under Nordic conditions. It has been revealed that Festuca species can contribute to increased drought resistance of Festulolium hybrids as compared with Lolium ssp. (Humphreys et al., 2013) and this was attributed to their increased root depth to access water more efficiently.

N fertilization levels had decisive effects on mixture composition

At N220, legume species had virtually disappeared already in the first harvest year (Fig. 1), in line with previous studies reporting adverse effects of N fertilization on legume proportion in grass–legume mixtures (Nassiri and Elgersma, 2002; Nyfeler et al., 2009). These high N levels applied to the grass–legume mixtures are far in excess of normal application rates for the area, which are between 120 and 150 kg N ha^–1 year^–1 for grass monocultures (Helgadóttir et al., 2014), and it can be inferred that the nutrient status of the grasses was not N-limited at the highest N level. The two grass species in the mixture start their growth early in spring and produce most of their biomass in the first half of the growing season, whereas the legume species are seasonally late developers. Thus, in a mixture with the legumes, the grass species took advantage from their competitive uptake for soil N (Nyfeler et al., 2011), produced high biomass yields early in the year and rapidly outcompeted their legume companions. Considering the comparably high yields in the grass-dominated mixtures at N220 (Supplementary data Fig. S2), competition for light might also have played a decisive role. Consequently, no grass–legume interactions could be estimated at these high N levels. Grass–grass interactions were not found either, probably because N was not a limiting growth factor and, thus, potential niche complementarity in N uptake due to for example differing rooting depths did not increase N uptake.

At N20 and N70, legume proportions were maintained in the mixtures over the whole experimental period and were, on average across the first 4 years, 51 and 27 % at N20 and N70, respectively (Fig. 1). This demonstrates that the legumes were able to survive at the lower N levels, most probably because they fully benefited from symbiotic N₂ fixation under N-limited conditions, thus having a relative advantage against grasses. The result agrees well with the conclusion that N acquisition through symbiosis is sustained even at these marginal environmental conditions (Rasmussen et al., 2013), although this conclusion needs to be confirmed by actually measuring symbiotic N₂ fixation for the two legumes in the current system.

Mixtures consistently suppressed weeds

Mixtures consistently suppressed weeds during the course of the experiment at all N levels and contained less biomass of unsown species than monocultures (Table 3; Supplementary Data Table S3), in accordance with previous studies of similar systems (Frankow-Lindberg, 2012; Suter et al., 2017; Connolly et al., 2018). Moreover, in the Agrodiversity Experiment, it has been shown that weed biomass was much less variable in mixtures compared with monocultures, and this was evident in mixtures with greatly varying species proportions, indicating that species evenness plays only a minor role in weed suppression (Connolly et al., 2018). The positive mixture effect on weed suppression in these studies may generally be attributed to more efficient resource use in mixtures than in monocultures (Suter et al., 2017).

Legume monocultures were severely infested with weeds already in the first harvest year at all N levels compared with the grass monocultures (Table 3; Supplementary Data Table S3). Other studies have shown that weed invasion is more pronounced in legume monocultures than in grass monocultures (Mwangi et al., 2007). Legumes generally increase N availability to unsown species, which probably increases weed root mass and density of roots, whereas grasses compete strongly for available N in the soil (Nyfeler et al., 2011; Hofer et al., 2017). Monocultures of T. pratense seem to suppress weeds well in the first harvest year (Frankow-Lindberg, 2012; Suter et al., 2017); however, in an experiment in Southern Sweden, weed biomass in T. pratense monocultures was strongly increased in the second year (Frankow-Lindberg, 2012).

Agronomic implications for marginal regions

In the northern parts of the Nordic countries, where agriculture is predominantly grassland based, the most common practice is to sow monocultures of P. pratense and apply 120–150 kg N ha^–1 year^–1 of mineral fertilizer. Forage legumes have only played a minor role so far because of their unpredictability within and between years. The primary argument for increasing the use of grass–legume mixtures in grassland systems is their potential to contribute to sustainable intensification of agriculture through savings in N fertilizers and an increase in resource efficiency (produce equally with less N input) and thus reduce emissions of reactive N forms to the environment (Lüscher et al., 2014). Our study clearly demonstrates that even in the less productive agricultural systems of the marginal areas, such grass–legume mixtures can be important contributors to reach this goal. This point is further accentuated by the fact that even though absolute yields are generally low in these marginal areas, the relative gain of mixtures compared with monocultures was substantial in the present study.

Even though higher total biomass yields were obtained from the grass monocultures at N220 compared with mixtures at lower N levels, this difference becomes less relevant when we consider the contribution of weeds to total yield, which made up around 25 % of the herbage in the grass monocultures at N220 but only around 10 % in the mixtures at both N20 and N70, on average over the first four harvest years. It can therefore be argued that growing grass–legume mixtures at moderate N levels has several advantages to growing grass monocultures at high N levels. First, comparable biomass yields were obtained from the two systems if we only consider the yield of sown species, resulting in clear economic benefits as mentioned above (same output from less input). Second, reduced application of mineral N fertilizers has considerable environmental benefits as it reduces energy use (Kitani et al., 1999), nitrate losses (Jensen et al., 2012) and greenhouse gas emissions (e.g. Schmeer et al., 2014). Third, the substantial legume proportion of mixtures at N20 and N70 in all years has a clear nutritional advantage. Numerous studies have demonstrated that legume-based forage can improve livestock performance compared with pure grass. The digestible matter in legumes is more easily accessible compared with that in grasses due to a higher proportion of cell content. This means that legume-based forage has a higher nutritive value, and the rate of decrease in quality with reproductive development is less rapid (Lüscher et al., 2014; Dumont et al., 2015). A recent meta-analysis has, for example, demonstrated that legumes in the forage improve dry matter intake and milk production as compared with pure grass diets (Johansen et al., 2018).

Conclusions

We observed robust diversity effects across a wide range of species composition over a period of 5 years in the current experiment. Positive grass–legume interactions suggest that symbiotic N₂ fixation is maintained even under the prevailing marginal growing conditions, provided that adapted species and cultivars are used. Our results therefore clearly show that even in the less productive agricultural systems, grass–legume mixtures can contribute to a more sustainable agriculture. However, maintaining legumes in the sward still remains a challenge in these environments. Suitable sward management strategies need to be found to stabilize the legume proportion, which may at first occur through adjusted N fertilizer application rates and harvesting regimes. It is also important to increase the persist- ence of legumes in the sward through breeding, and thus further improve the robustness of the system.

SUPPLEMENTARY DATA

Supplementary data are available online at https://academic.oup.com/aob and consist of the following. Appendix A: supporting text and information on the methods and analyses. Table S1: sowing proportions in monocultures and mixtures of the four forage species. Table S2: regression coefficients for the effects of species proportions and N fertilization on total yield (including unsown species). Table S3: observed biomass percentages of unsown species in the four monocultures. Figure S1: observed vs. predicted biomass yield based on multiple linear regression for each of the five experimental years and averaged across the 5 years. Figure S2: observed total yields (including unsown species) of pooled grass monocultures and mixtures for the five experimental years at the highest N fertilization level (220 kg N ha⁻¹ year⁻¹). Figure S3: predicted total yield averaged across 5 years as a function of legume proportion omitting the parameter θ.