Farm sample and sampling scheme
We sampled farms according to several criteria. First, eligible farms had two commercial livestock production enterprises, i.e., beef cattle associated with either meat sheep, pigs, or poultry. Second, the whole farm was managed according to OF principles and had completed the conventional-to-organic conversion period. Third, at least a half full-time equivalent (FTE) worker was managing the system to ensure that the farms sampled were viable enterprises.
Once these three criteria were met, we opted to sample grassland farms at different altitudes (in lowlands and uplands) with contrasting marketing strategies (direct sales, integrated supply chain, etc.) and with or without crops or on-farm processing (meat-cutting workshop, mill, etc.). Farms in the sample sometimes counted a third animal species (e.g., horses or backyard chickens), but not for economic purposes.
Farms were surveyed during the MixEnable project (https://projects.au.dk/coreorganiccofund/core-organic-cofund-projects/mix-enable/). We consulted the administrative and advisory services of the Auvergne, Limousin, and Occitanie regions to obtain a list of farms meeting these criteria. Sixteen farmers agreed to be interviewed. An additional farm is the Salamix farmlet experiment (https://www6.inrae.fr/experimentations-systeme/Les-experimentations/Elevage/Mixte/Salamix), which is a beef cattle–sheep grazing system managed by a half FTE worker at the INRAE Herbipôle research facility in Massif Central uplands (45° 39′ 02.9 N 2° 44′ 01.0 E).
In the final sample of 17 organic farms, beef cattle were associated with meat sheep (n = 7), pigs (n = 6), or poultry (n = 4), with each species representing at least 10% of farm livestock units (LUs). Three farms with direct sales were below this species-ratio threshold but were retained to increase the diversity of the sampled MSL systems. The LU concept is widely employed in livestock farming system analyses to quantify the farm stocking density. It is based on the same shared unit among species and types of animal products but does not account for within-species variability in animal size and feeding requirements. Hence, for herbivores, we adjusted the LU coefficients according to cow and ewe metabolic weight to account for breed differences, as metabolic weight is the variable used to calculate maintenance requirements and additional requirements for production and draught power (IPCC 2019). We therefore multiplied the LU coefficient of each animal category by the ratio of the metabolic weight of the dam to the metabolic weight of a baseline dam counted as one animal unit (Smith et al. 2017). In France, one LU corresponds to a-600 kg liveweight dairy cow producing 3000 kg of milk and eating 3000 feed units (FU) per year (where 1 FU is the energy content of 1 kg of barley; INRA 2018) or 4500 kg of dry matter for forages (Institut National de Gestion et d’Economie Rurale 1989). Based on this information and using an equivalence based on metabolic weight, a standard meat ewe weight of 52 kg corresponds to 0.15 LU. Monogastric LU coefficients were adapted based on energy and thus concentrate consumption in OF with the help of experts (Antoine Roinsard, ITAB, France and Marie Moerman, CRA-W, Belgium). We considered that an organic finishing pig consumes an average of 420 kg of concentrate (1 kg concentrate = 1 FU) regardless of the age at slaughter; therefore, we used a coefficient of 0.14 LU (0.14 = 420/3000) for each pig sold. An organic sow consumes an average of 1500 kg of concentrate in 1 year, and all types of systems combined correspond to a coefficient of 0.5 LU (0.5 = 1500/3000). Replacement sows were counted for 0.14 LU and piglets for 0.055 LU (Cohen and Zahm 2011). Organic laying hens consume an average of 44.5 kg of concentrate per year, corresponding to 0.014 LU (0.014 = 44.5/3000), while organic broilers consume an average of 6.2 kg of concentrates regardless of the length of the rearing period, which corresponds to 0.002 LU (0.002 = 6.2/3000) for each broiler.
Data collection
The survey was built to obtain the data needed to obtain an overview of the farm structure, management, and performance in a reasonable time, i.e., no more than 3–4 h for the whole survey (a necessary condition for farmers to accept to receive us). It was composed of eight parts: farm structure (i.e., agricultural area and numbers of full-time and part-time workers), livestock, pastures, crops, sales, input purchased and use of byproducts, economics, and farmers’ perceptions of their work. Each farmer was visited once for the 3–4-h interview, and many questions assessing quantitative or binary data were asked. Open-ended questions were also asked on practices related to grazing and pasture, effluent, and crop management. Farmers were also asked to provide (for later office analyses) the farm’s accounts to calculate economic and quantitative indicators of sales and purchased inputs. Economic data were recorded for 2017, which is considered an average year in terms of climate and market conditions. Only 13 of the 16 commercial farmers agreed to share their farm’s accounts; therefore, the economic assessment was not performed for three farms or for the INRAE farmlet experiment.
Farm operating analysis integrating the ecological network analysis
We used ENA, a holistic approach, to describe, quantify and analyze interactions among farm components. First developed for econometrics, Hannon (1973) applied this input-output analysis to ecology to describe the structure of the ecosystem and quantify within-system relationships. Rufino et al. (2009) applied a network analysis to “quantify the degree of integration and diversity of farm household systems using a set of indicators.” Stark et al. subsequently proposed a new way of understanding and characterizing complex systems by considering the interactions between system components using indicators derived from ENA and used this method to analyze the benefits of crop-livestock integration in Latino-Caribbean farms (Stark et al. 2016, 2018). Here, we adapt this conceptual model to the case of MSL farms in temperate areas. Model implementation involved two stages: conceptualization and modeling.
The conceptualization stage consisted of defining system boundaries — in space and time — and components and identifying interactions among system components and between the system and its environment. Here, we worked at the farm level and adapted the segmentation of system components to fit the study objectives. Two animal components (beef cattle plus either meat sheep, pigs, or poultry) accounted for livestock, possibly with a third not-for-profit animal component being added to some farms. Annual grain crops and forage crops (such as corn silage or forage meslin) were aggregated into a single crop component. Grasslands were subdivided into two components, i.e., permanent (PG) or temporary (TG) grasslands, as each has its own management practices and a specific role in coupling carbon and N cycles. Consistent with Stark’s model (Stark et al. 2016), we represented effluent and food storage as two distinct components. This distinction provided a link to the conceptual framework for levels of crop-livestock integration proposed by Bell and Moore (2012). We represented direct and indirect flows between components. For example, between grasslands and animals, direct flows occurred during grazing (grass intake and dejection) and were collocated interactions. Indirect flows relied on stored fodder or manure flows and were segregated interactions between animals and land. For example, animal dejections indoor were stored and then allocated to PG, TG, or crops according to farmer strategy. We split the fodder and grain storage sites to represent fodder flows from catch crops or cover crops separately from grain flows. In the present study, a cover crop was defined as a fast-growing crop that was grown between successive sowings of a main crop. It could be destroyed and left on the field, grazed by livestock or harvested for silage. One or more processing components, such as a meat-cutting plant or a mill, were added to some farms. Inputs corresponded to all the biomass entering the farm gate, i.e., feed, animals, organic fertilizers, seeds, and manure, whether purchased, received, or exchanged with neighbors. Input flows came from outside the system and arrived directly to the component where they were used. Outputs included sold, self-consumed, or exchanged farm products. Plant and effluent outflows left the system from their respective storage component, while animal outflow left the system from the animal component. On-farm processed products left the system from the processing component. Interactions corresponded to biomass flows between components or with the environment, and they reflected management practices such as feeding, harvesting, and manuring. The system runs over a production cycle of 1 year, and thus we assessed all the flows (purchases, sales, agricultural practices) and performances on this yearly basis.
Modeling consisted of selecting a unit and quantifying flows and storage changes. Flows were reported annually based on the quantity of biomass exchanged and multiplied by the N content (Table 1). We chose N as a common unit because N is an essential and mostly limiting nutrient in agriculture (Rufino et al. 2009). Moreover, N is viewed as a resource supporting production and as an environmental burden. Some biomass flows were not recorded on-farm and had to be estimated, such as grass intake and dung and urine excretions. Forage intake from cover crops or crop residues was estimated according to theoretical animal requirements of 4500 kg of dry matter per LU per year for the standard dairy cow (Institut National de Gestion et d’Economie Rurale 1989), number of grazing days, and herd size. Theoretical animal requirements were used, assuming that the needs of the animals were covered. Grass intake on pastures was estimated on a daily basis from animal requirements minus the consumption of hay and silage and estimated forage intake of cover crops and crop residues.
Table 1 Principal nitrogen coefficients used in the analysis For animal components, the total amount of N excreted was calculated as the difference between N inputs (feed and animals purchased) and N outputs (sales or self-consumption). Effluent-related gaseous emissions were deducted from the N excreted to estimate the amount of N reaching the ground. These effluent-related gaseous emissions were calculated using the method reported by Gac et al. (2007), accounting for farming practices and the type of manure. Nitrogen that is not lost is assumed to be applied on pastures or crops. Plowing a TG makes N available for the next crop. The amount of N released after plowing grassland depends on the nature of the soil, age of the grassland, and grazing practice (Davies et al. 2001). However, as this accurate information was not available from the surveys, we worked with a fixed value of 100 kg N.ha−1 (Comifer 2013) that was multiplied by the average area of TG turned into cropland area each year. As grasslands were separated into TG and PG, we allocated forage production (grazed and harvested) to these two categories. This information was unable to be collected in the surveys, as the farmers were unable to provide a detailed grazing schedule. We thus proposed an approach that allocates fodder production to the TG and PG according to the management method. This approach required a series of assumptions: (1) a grassland that was only mowed was qualified as TG, as it was probably a pure legume or very productive grassland; (2) “harvested then grazed” grassland area was compared to the TG area. If the former was higher than the latter, then we considered that all TGs were “harvested then grazed,” and some PGs were also managed in this manner to obtain the hectares of “harvested then grazed” grasslands. If the “harvested then grazed” grassland area was less than TG area, then the “harvested then grazed” grassland area was considered as TG; (3) a grassland that was only grazed was considered as non mechanizable and therefore as PG; and (4) wrapped-bale and grass silage were produced by mowing TG. These assumptions allowed us to calculate the relative share of hay, for example, area from TG and PG. Knowing the total quantity of forage (hay, silage, and wrapped) produced over the year and the distribution ratios between TG and PG enabled us to allocate the quantities of harvested forage to PG and TG components. We then calculated a “corrected fodder stocking rate,” i.e., the number of herbivore LUs per fodder area (grassland + forage area) corrected by the number of LU equivalents fed with fodder from crops (intercropped and covered) and fodder purchased off-farm. This corrected fodder stocking rate allowed us to estimate a potential yield for grassland. We considered a similar potential yield for both TG and PG, as (i) PGs were productive mesophile grasslands and (ii) no additional information was provided by farmers. This potential yield enabled us to calculate the total potential production in tons of dry matter for each component. We then deducted the respective quantities of forage harvested to estimate the amount of grass potentially grazed in TG and PG.
N losses (run-off, leaching, and volatilization) largely depend on soil type, climate, and plot management. Accounting for these factors would require many approximations. For instance, the soil type cannot be precisely characterized, as some of the farms are located in areas where the soil type is not registered on national soil maps, while different soil types are also present on the same farm. We therefore did not quantify N losses at the farm level. N symbiotic fixation and N deposition were also not considered in the farm operation model to ensure a consistent input-output approach, as described in the studies by Rufino et al. (2009) and Stark et al. (2018). As some systems were not at equilibrium, inventory changes (end stock–beginning stock) were considered by adding positive changes to exports and negative changes to imports.
Matrix construction and selected ENA indicators
A matrix was then created with rows as the origin of N flows (0 to n components) and columns as the destination of N flows (1 to n+2 components), where n is the number of system components. The value of the flow reported in kg of N per hectare of usable agricultural area (UAA) per year is at the intersection. Row 0 corresponds to the inflows, i.e., imports (purchases). The last column corresponds to nonvaluable outflows, such as animal losses and gaseous emissions, and the penultimate column corresponds to valuable outflows sold, exchanged with neighbors, or self-consumed.
Five network analysis indicators were calculated from this matrix, as described below. System activity corresponds to the total system throughflow (TST), which represents the sum of all flows circulating throughout the system, including both inflows and throughflows (Finn 1980; see Fig. 2). Total internal throughflow (TT) is the sum of all flows circulating among components, which represents the interaction activity. Link density (LD) is the ratio of the number of internal flows to the number of components (Latham 2006) and represents the internal flow density. The average mutual information (AMI) and the statistical uncertainty (Hr) are used to assess the level of organization of the network of flows (Rutledge et al. 1976; Rufino et al. 2009; Stark et al. 2018). The AMI quantifies the organization of the flow network, where Hr is its upper bound. Hr illustrates the diversity of possible flows given the size of the system and the amount of throughput (T..), corresponding to the total quantity of N from all inflows, outflows, and flows within the system. If the total flow in a system is divided equally between all components and all components are connected, then AMI will be zero or very close to zero. Conversely, if the flow network is unbalanced with only a few main flows connecting few components, the AMI value will be close to its upper boundary (Hr). The AMI/Hr ratio is the proportion of diversity achieved based on the possible connections between system components (Rufino et al. 2009).
Both AMI and Hr consider the inflows and internal flows in their calculations. Here, we proposed a new indicator inspired by AMI/Hr that only considered the internal flows to strictly assess internal flow organization (Eqs. 1 and 2). We defined the internal flow organization indicator as 1 – AMIintern/Hrintern. The value of this indicator approaches 1 when flows are equally distributed among all components, indicating that the flow distribution is homogenous and approaches 0 when the flow distribution is heterogeneous. Tij is the value of the flow from component i to component j, where i is the number of rows and j is the number of columns. Ti. is the total outflow for component i, and T.j is the total inflow for component j:
$$ {AMI}_{intern}=\sum \limits_{i=1}^n\sum \limits_{j=1}^n\frac{T_{ij}}{T_{..}}\ {\mathit{\log}}_2\frac{T_{ij}{T}_{..}}{T_{i.}{T}_{.j}} $$
(1)
$$ {Hr}_{intern}=-\sum \limits_{i=1}^n\frac{T_{i.}}{T_{..}}\ {\mathit{\log}}_2\frac{T_{i.}}{T_{..}} $$
(2)
All flows were calculated using Excel to build a matrix. One file was created per farm. The indicators were calculated with R software using the enaR package (Lau et al. 2017). We adapted the calculation code for the organization indicator included in the R package to account only for internal flows and thus calculated the internal flow organization indicator.
Farm structural indicators
Farm operations mainly depend on farm structure factors, i.e., farm area, different workers, herd size, and composition. Moreover, the farm structure can influence indicator values (TT and TST), as farm size influences herd management and the technical efficiency of the production system (Veysset et al. 2015). The level of crop-livestock integration also affects farm performance (Ryschawy et al. 2012; Lemaire et al. 2014). We therefore included structural indicators in the analysis. We used UAA in hectares, the total number of workers in FTE and the total number of LUs to characterize the size of the farm. The percentage of crops (%crops/ha), i.e., annual crops, including fodder crops, and the percentage of PG (%permanent_grassland/ha) within the UAA, provided additional information on the farm structure. The percentage of beef cattle LUs among total LUs (%beef_cattle/LU) on the farm provided information on the level of species mixing. The “fodder stocking rate,” calculated using the method described in Section 2.3 without the correction, was used to evaluate the level of intensification of the fodder area. The “global stocking rate,” defined as the total (monogastric and herbivore) number of LUs per hectare of UAA assessed the level of animal density at the farm level.
Farm performance indicators
We assessed the multiperformance of MSL farms by accounting for the three dimensions of sustainability. We retained three criteria: resource-use efficiency, environmental impact, and human well-being. The efficiency analysis highlights farms that are able to produce a large amount of output with little use of inputs or a moderate output with no inputs, thus defining low-input systems. Resource-use efficiency was assessed under N, economic, and net protein efficiency (NPE) conditions. N efficiency, which was calculated from the ENA matrix, was defined as the ratio of N output to N input. Economic efficiency was defined as the ratio of added value to gross farm output (including subsidies in both variables). Added value corresponded to the gross farm output (sum of products sold and of stock variations minus purchased animals) minus intermediate consumption (i.e., the sum of goods and services purchased that are required to produce) and depreciation cost of equipment. Ruminants receive CAP subsidies that are essential to achieve profitability. We therefore integrated subsidies into the calculation of added value to compare the economic efficiency of different farm types. Economic efficiency was calculated from the farms’ accountancy data. We also calculated farm NPE, i.e., the ratio of human-edible proteins contained in the food products to human-edible proteins in the feed consumed by the animals, using the corresponding human-edible protein content for each product proposed by Laisse et al. (2018). NPE facilitates an assessment of feed–food competition. A farm is a net producer of human-edible protein if it produces more food protein than it uses feed protein, i.e., if the ratio value of both variables reported in kg of edible protein is higher than one.
Environmental impact was assessed through N balances. We applied the Economic Input:Output (EIO) budget to calculate the N balance and the Biological Input:Output (BIO) budget to calculate the N balance with biological N fixation (BNF) (Watson and Atkinson 1999). Both parameters were calculated at the farm scale and accounted for N sales and N purchases over the farm gate. N sales corresponded to N outputs in the ENA matrix and N purchases to N inputs. In the N balance with BNF, the budget includes inputs from symbiotic N fixation and atmospheric N deposition, which enables the integration of the potential of legumes for closing the N cycle, which is especially important in OFs where chemical inputs are banned. BNF was calculated from the equation provided in the study by Anglade et al. (2015) applied to harvested yields, which includes belowground contributions such as N cycling associated with roots, nodules, and rhizodeposition. For the cereal legume mixture, the percentage of legumes in the crop was set to 20% (Vertès et al. 2015). For grasslands, we set a specific legume content for each type of grassland as follows: 13.3% for TG and 11% for PG (INRA 2018). The yield of a meadow, which was used to assess N export, corresponded to the sum of the grazed and harvested amount of grass. Field N deposition is a surface-dependent parameter. For this study, we used a single value of 10 kg N ha−1 year−1 (Dentener et al. 2006), which was multiplied by the UAA of each farm.
Farmer well-being at work was addressed by measuring three indicators: physical difficulty of work, mental workload, and overall satisfaction. The OECD also uses subjective indicators of well-being to analyze the impacts of their policies on the population (Boarini et al. 2012). Here, physical difficulty and mental workload were assessed because work is often considered more difficult physically in MSL systems due to the limited mechanization options (Martin et al. 2020). Labor is also quantitatively different from that in specialized systems with different types of knowledge, skills, and farmer capacities for monitoring system performance that can increase the mental workload (Kingwell 2011). In cattle-sheep systems, farmers also mentioned the pleasure of varied work and flexibility of the work organization as a matter of satisfaction (Mugnier et al. 2020). The assessment of these three indicators was based on farmers’ perceptions, which is commonly used to evaluate farmer well-being (Besser and Mann 2015). During the surveys, farmers were asked three questions, one on each indicator, and farmers’ perception indicators were scored on a 4-point scale, with a score of 1 indicating less satisfied and a score of 4 indicating most satisfied.
Analysis of the indicators
After performing univariate and bivariate analyses, a principal component analysis (PCA) was carried out on thirteen active variables (eight structural variables and five ENA variables) (Table 2) using R software and the FactoMineR package (Lê et al. 2008). Performance indicators were considered illustrative variables, and thus they were not included in the construction of principal components. We then conducted a hierarchical clustering on principal components (HCPC) analysis of the four first factors, again using the FactoMineR package. From the HCPC results, we calculated the ratio of intra inertia to total inertia, which illustrates the percentage of data variability in the HCPC analysis. We conducted a nonparametric test of multiple pairwise comparisons (Dunn) to determine whether the clusters performed identically in terms of the distributions and median values. The significance threshold was set to a p value of 5%. Each cluster was illustrated by its paragon (individual closest to the barycenter of the cluster) using the R package igraph (Ognyanova 2016) to represent the flows between system components and between components and the environment. The thickness of the arrow illustrates the intensity of the flow. The number of arrows according to the number of components illustrates the density of flow. Flow organization was defined as homogeneous when all the arrows had a similar thickness. When arrows of different thicknesses were observed, the flow organization was considered heterogeneous.
Table 2 Farm operating and structural indicators calculated for the seventeen farms studied