A method to estimate the environmental impacts from genetic change in pig production systems

2.1 Goal and scope

The goal of this study was to investigate how GTs in pigs can be included in a livestock LCA framework. Further, we wanted to explore which GTs influenced the environmental impacts of the system and how their interactions could affect future breeding goals in pigs. The study boundaries were cradle to farm gate and the functional unit was 1 kg of live animal at the farm gate (FAO 2016). The functional unit was chosen to avoid uncertain assumptions on the killing out percentage and the meat quality which are certainly affected by traits but are difficult to estimate from animal modelling. The model used attributional LCA (BSI 2011) and distributed impacts among coproducts by economic allocation (Mackenzie et al. 2017).

All EI categories from the ReCiPe 2016 midpoint method (Huijbregts et al. 2016) were considered in this study. To align with FAO (2018), all impact categories with sufficient available data were calculated and the most important impacts are reported and discussed in this paper: global warming potential (GWP), terrestrial acidification potential (TAP), freshwater eutrophication potential (FEP), land use (LU), and fossil resource scarcity (FRS).¹ Available impact categories that are not commonly reported in animal LCA studies are reported in the Electronic Supplementary Material - ESM, section 8.13: Stratospheric Ozone Depletion, Ionizing Radiation and Mineral Resource Scarcity. Due to data unavailability, ozone formation, fine particulate matter, toxicity, and water consumption impacts were not calculated.

2.2 Choice of system

An intensive indoor pig production system was chosen since most of the pigs produced in Europe are raised as such (Driver 2017). The Danish pig production industry has published abundant information which, together with the influential Danish contribution to the European pig semen and pork market, gave an opportunity for detailed modelling of a highly defined, representative system that is recognized as an international benchmark. To evaluate a more holistic perspective of the influence of GTs on the full life cycle of the pig, an integrated pig production system, which contains both pig breeding and a production unit, was chosen. In this paper, the Baseline System will be used for the unmodified system where all variables have their mean value, and the Baseline System is therefore deterministic.

2.3 The system

The model covered 3 stages in the life of the animals: (1) gilt: female pigs from weaning until the point of mating, where the gilts were fed together with the production pigs until day 56 and thereafter fed a specialist gilt diet; (2) sow: female pigs from first insemination until culling including piglets until weaning; this stage includes gestation, lactation, wean to oestrus, and failed pregnancies; (3) production pig: growing and finishing of pigs from weaning to slaughter weight at farm-gate (see Fig. 1). The terms gilt, sow, and production pig will be used in the following text to refer to these three stages. Male reproducing pigs were not included in the model since preliminary calculations estimated that thousands of offspring may result from a single boar (> 3000 and > 95,000 for natural and artificial insemination, respectively) and their influence on the slaughtered pigs could, therefore, be expected to be minimal. Time spent in each stage, diet composition, and feed amount for the Baseline System can be found in the ESM, section 8.1–8.3.

Open image in new window

The LCA model was constructed to simulate the integrated Danish intensive indoor pig system based on the method described by Mackenzie et al. (2015) (see Fig. 1). The structure of the system was based on recommendations for pig management published on the home page of the national Danish Pig Board (SEGES 2018). A sow compartment, containing the first two life stages, following the sow from birth to 8th parity, was constructed, with feed, manure, and litter sizes adjusted according to age and surviving proportion of sows (calculated from phase specific mortality rates) (see section 8.1–8.3). A production pig compartment, equivalent to the third life stage, followed the pig from birth to slaughter at 110 kg, with feed intake and output of live pig adjusted for mortality rates. A single feed was chosen from weaning to slaughter for the production pigs since phase feeding is not commonly applied in Denmark (Sloth 2000) and trials on phase feeding have previously shown reduced leanness in Danish trials (SEGES 2013a). Diet was composed of typical commercial feed ingredients, with 6 different feeds (creep feed, production pig feed, gilt feed, early gestation feed, late gestation feed, lactation feed) offered to the sow and 2 different feeds (creep feed, production pig feed) offered to the growing pigs, according to published information and the recommendations for the Danish pig industry. The composition of all feeds can be seen in the ESM, section 8.1. Emissions from feed ingredients and power consumption on farm were taken from inventories for Denmark or similar countries from the SimaPro 8.5.2.0 database (Pre Consultants 2017). Manure chemical composition was calculated based on the mass balance principle, and emissions from the pig unit and manure management were calculated based on (IPCC 2006) for GWP and (Guinee 2002) for other chemicals.

2.4 Pig traits considered in the LCA model and their origin

2.5 Energy balance model

The pig model specified inputs and outputs as functions of GT, and feed intake was based on energy demand from growth and maintenance. Feed intake was calculated from the required energy for each day based on (Wellock et al. 2003; Dourmad et al. 2008; Tallentire et al. 2016) with daily energy intake calculated as

$$ {E}_f\times {R}_d={E}_m+{E}_g $$

(1)

where E_f is the metabolizable energy in the feed eaten each day, R_d is the energy use efficiency which relates apparent energy use to available energy, E_m is the energy needed for maintenance of the body, and E_g is the energy needed for growth. By dividing Eq. (1) with R_d and expanding the right side of the equation, we get:

$$ {E}_f=\frac{1}{R_d}\times \left(a\times {BW}^b+\left(\frac{\Delta \mathrm{Prot}\times 23.8\raisebox{1ex}{$\mathrm{MJ}$}\!\left/ \!\raisebox{-1ex}{$\mathrm{kg}$}\right.}{0.6}+\frac{\Delta \mathrm{Lip}\times 39.6\raisebox{1ex}{$\mathrm{MJ}$}\!\left/ \!\raisebox{-1ex}{$\mathrm{kg}$}\right.}{0.8}\right)\right) $$

(2)

where a is the linear metabolic constant corresponding to Maintenances, BW is the body weight at the start of the day, b is the metabolic maintenance exponent, ΔProt is the growth of protein, and ΔLip is the growth of lipid. R_d was estimated for each feed with the “feed mix nutritional calculator” tool published by SEGES (2017) and the values of a and b were taken from van Milgen et al. (2008). It was chosen to model the E_m based on the metabolic body mass over the protein mass since previous published and grey literature reported the body mass of the pig much more frequent than the protein. ΔProt was estimated from nitrogen content in pigs slaughtered at different ages and transformed to crude protein by multiplying with 6.25. ΔLip was estimated by dividing ΔProt with ProtLip, which was based on energy conservation in a simplified version of Wellock et al. (2004) updated to present pig performance. The E_m was calculated for each day in the gilt and production pig stages with fast growth, accounting for the daily BW by assuming linear growth within a stage, and calculated for the average weight multiplied by the number of days for the sow stage with slower growth (see ESM, section 8.11 for energy implementation of traits). The LCA model was adapted on matrix form according to Heijungs and Suh (2002) (see Fig. 2):

$$ \boldsymbol{B}\times {\boldsymbol{A}}^{-\mathbf{1}}\times \boldsymbol{f}=\boldsymbol{g} $$

(3)

where B is the environmental matrix, containing impacts for each process, A is the technology matrix, containing the flow of resources inside the model, f is the functional unit, and g is the result matrix. The model was built in MATLAB R2017a (MathWorks 2017) with the B matrix imported from single inventories from SimaPro 8.5.2.0 (Pre Consultants 2017). A more thorough description of the matrix adaptation can be found in the ESM, section 8.9.

Open image in new window

2.6 Used data and compilation of correlation matrix

Genetic variance and significant genetic correlations for pig traits were collected from 22 recently published articles on genetic variance and correlations in pigs (see ESM, section 8.12). The selected studies contained genetic correlations, additive genetic variance, and/or data which made it possible to calculate either of these under the assumption of additive genetic effects. Only studies conducted on common and commercially available modern fast-growing pig breeds raised indoor in Europe, North America, or Australia were included. Studies that did not list the number of pigs sampled were excluded. If a study investigated multiple breeds, each breed was included in the sample. If a study presented multiple results from different models, only results from the model the study recommended was included in the sample.

Trait correlations with measured or obvious high correlations were merged. Average daily feed intake was in this way replaced with ADG100, age at first farrowing was replaced with AgeMature, feed conversion ratio was replaced with residual feed intake (later transformed into Maintenance, see ESM, section 8.5), litter size and total litter weight for any parity number were replaced with LitterSize, and sow life time and sow lifetime piglets produced were replaced with AMPar. The correlation matrix was pooled with univariate weighting applying Fisher’s z transformation to reduce bias with different sample sizes and correlations (Furlow and Beretvas 2005). The variance was pooled from the studies weighted with sample size applying Bessel’s correction (Upton and Cook 2014). If there was no known genetic correlation between any two traits, zero correlation was assumed. Since the pooled correlation matrix was not positive semidefinite, a small correction to the matrix was implemented to the nearest semidefinite matrix (Higham 2002). It was assumed that all GTs were normally distributed.

2.7 Uncertainty and sensitivity

To test the range of outputs from the model and to evaluate the effect of every included trait in the model, a number of uncertainty and sensitivity analyses were performed. This article follows the proposed definition by Saltelli et al. (2008), where the term uncertainty analysis (UA) comprises estimation of the variance of the model outcome and sensitivity analysis (SA) refers to estimation of relative contributions of individual variables to the overall variance.

3.1 Outcomes of the LCA model

The production pig component contributed the majority of the environmental impacts, followed by the sow and the gilt system. TAP and FRS were more dominated by the sow component than the other impacts and LU had the highest contributions from the production pig system.

3.2 One-at-a-time sensitivity analysis

The results of the OAT are shown in Fig. 3. All environmental impacts were sensitive to maintenance (12.8%, 18.8%, 11.5%, 12.6% and 9.6%/− 12.5%, − 18.4%, − 11.2%, − 12.4%, and − 9.2% for GWP, TAP, FEP, LU, and FRS in + 2SD/− 2SD direction, respectively) with ADG100 being the second most influential trait for most EI in the − 2 SD change (2.9%, 3.6%, 3.2% 2.9%, and 2.8% for GWP, TAP, FEP, LU, and FRS respectively in – 2 SD direction); BWMature, BWLossLactation, LitterSize, ProtLip, and AMPar had some influence (around ± 2%), whereas ADG30, AgeMature, LactationFeed, PreWMort, and PostWMort had minimal influence on the model outcome (< ± 1%). ADG100, ADG30, BWlossLactation, ProtLip, and LitterSize reduced impacts when they were increased, whereas BWmaturity, Maintenance, and AMPar increased impacts when they were increased. TAP was the most sensitive environmental impact to trait variation; GWP, FEP, LU, and FRS had similar sensitivity to variation of traits among the EI for most traits.

Open image in new window

3.3 Correlated cluster sensitivity analysis

The results of the COT can be seen in Fig. 4. The figure shows that when all traits in a cluster were changed by ± 2 SD, cluster 1 resulted in the highest change in environmental impacts, cluster 2 had an intermediate effect, and cluster 3 resulted in nearly no change in environmental impacts compared with the Baseline System. For clusters 1 and 2, it is clear that TAP was the most sensitive environmental impact to changes in all clusters. When all traits in cluster 1 were reduced by 2 SD according to their correlations (left side of Fig. 4), all environmental impacts were reduced. By reducing the cluster 1 traits in Fig. 4 by 2 SD according to their internal correlations, the TAP was reduced around 17% and other impacts were reduced around 10% relative to the Baseline System. Cluster 2 had reduced environmental impacts when all traits in the cluster were increased according to their internal correlations (right side of Fig. 4). By increasing cluster 2 traits in Fig. 4 by 2 SD according to their internal correlations, the TAP was reduced around 6% and other impacts were reduced around 4% compared with the Baseline System. Cluster 3 resulted in almost no change in EIs when the traits were changed, but had slightly reduced EIs when the traits in the cluster were increased according to their internal correlations (right side of Fig. 4). The EIs was slightly reduced by around 2% when the cluster 3 traits were increased by two SD.

Open image in new window

3.4 Sensitivity indices

The total SI, uncorrelated SI, and correlated SI, produced with 20,000 samples, can be seen in Fig. 5. All EI categories showed similar SI for most traits, although TAP and LU showed slightly higher SI in dominant traits and FRS showed slightly lower SI than GWP and FEP for dominant traits. ProtLip (0.39, 0.42, 0.40, 0.44, and 0.33 total SI for GWP, TAP, FEP, LU, and FRS, respectively) and Maintenance (0.83, 0.88, 0.84, 0.87, and 0.74 total SI for GWP, TAP, FEP, LU, and FRS, respectively) were the traits with the highest total SI in all EIs, and were thereby responsible for most of the variance in the model outcome. BWMaturity, BWLossLactation, and LitterSize had a total SI of a few hundredths. The correlated SI contributed about half of the total SI for Maintenance and almost all of the total SI for ProtLip, so interactions for these traits with other traits were very important for explaining the variance in the model. BWLossLactation and LitterSize had nearly all their total SI from the correlated SI, and ADG100 had slightly negative correlated SI. Maintenance was the only trait with a major uncorrelated SI although ADG100 and BWMaturity also had minor contributions from the uncorrelated SI.

Open image in new window

4.1 Outcomes of the LCA model

The environmental impacts from the baseline model in this study are consistent with literature values (Halberg et al. 2010; Nguyen et al. 2011). The result of this model is presented as a mean and standard deviation of 100 correlated sample runs and the SI was calculated based on 20,000 samples. As we have illustrated with the mean of variance and variance of mean, the correct sample size was essential to achieve accurate results without wasting excessive resources (Baldini et al. 2017; Mendoza Beltran et al. 2018). It was confirmed that the majority of all impacts comes from the production pig system, so pig LCA models only focusing on wean to slaughter is a reasonable approximation for the system boundary (Monteiro et al. 2016), but this assumption could underestimate some impact categories for the full production system with up to 29%.

4.2 Model sensitivity analysis

The contrasting results of the different sensitivity analyses demonstrated the importance of the methodological approach when drawing conclusions regarding the potential of GTs to reduce environmental impacts. The results of the OAT sensitivity analysis showed the importance of Maintenance, ADG100, and some sow performance traits (see Fig. 3), whereas Maintenance and ProtLip were highlighted by the SI methodology (see Fig. 5) as good options to reduce environmental impacts. Both outcomes align with the focus of the breeding industry on growth rate and feed efficiency (PIC 2017; DANBRED 2018). However, when evaluating the potential of current breeding strategies to reduce the environmental impacts of pig production systems, it is critical to know which traits are improved when selecting for feed efficiency. This is important because some traits may have biological limits, for example lipid content in the body cannot go below a theoretical zero, and this needs to be recognized if such limits are being approached (Tallentire et al. 2018).

The OAT results did not suggest any environmental benefits from improving sow longevity and reducing pig mortality, since the OAT showed little sensitivity to changes in these traits, while AMPar actually increased environmental impacts when increased. The low sensitivity of environmental impacts to LitterSize and AMPar could be due to the high influence of management for outcomes in this phase of the production system, such as the choice of when to cull a sow (Nagyné-Kiszlinger et al. 2013; Thekkoot et al. 2016). Further, only traits which affect major inputs and outputs of the LCA model, i.e. feed inputs, live weight pigs produced, and manure from pig systems (Mackenzie et al. 2015; McAuliffe et al. 2016), will have a major influence on the environmental impacts. Maintenance and ProtLip could therefore be expected to be influential since they affect feed intake for both the gilt, the sow, and the production pig. Traits like LitterSize and AMPar only directly influence the breeding phase of the production cycle for pigs when considered in an LCA model. Increases in AMPar increased environmental impacts because the higher BW as the sow increases in age requires higher feed intake without equivalent associated benefits in increased outputs, though this effect is small compared with other traits. Piglet mortality only has minor influence on the environmental impacts, since the invested feed to produce them is small compared with the full grown production pig. In a similar way, traits which only affects a short period in the life cycle (such as AgeMature and ADG30) or traits with only a small variance compared with the mean value (such as LactationFeed) can be expected to only have minor contribution to the outcome of the model.

Considering the genetic correlations in relation to the COT sensitivity analysis (Fig. 4), some classic breeding objectives may not be as effective as would be expected in reducing environmental impacts. For example, the COT suggested that, due to the negative correlations between on one hand ADG30 and ADG100 and on the other hand ProtLip, breeding for increased ProtLip is a much more effective strategy to reduce environmental impacts than increasing growth rate (see Table 3 and Fig. 4 cluster 1). Similarly, an increase in piglet mortality only has a minor effect on environmental impacts when compared with reductions in sow feed intake and higher litter size due to negative correlations between the traits in cluster 2. This could be due to some simplifications in the modelling process, where the interaction between sow and piglet performance was only included in the sampling process, and not as a causal effect implemented with deterministic equations. During the last decade, the pig production industry has moved towards higher sow replacement rates (AHDB Pork, SEGES, personal communication), though multiple publications have argued for lower sow replacement for economic and welfare reasons (Stalder et al. 2003; Onteru et al. 2011; Gruhot et al. 2017). In this study, sow longevity and its interaction with AgeMature had very little influence on the environmental impact categories tested for cluster 3, which could be partly due to their negative correlation.

The total SI highlighted the sensitivity of the environmental impacts to changes in Maintenance and ProtLip; the sensitivity to other traits, including ADG100, was much lower. In a similar pig LCA study on correlated traits (Reckmann and Krieter 2015) also found that the environmental impacts of pig production systems were most sensitive to ProtLip but showed low sensitive to ADG100. ProtLip was one of the most, but ADG100 was one of the least, sensitive traits. This is in good agreement with the total SI where ProtLip was very sensitive and ADG100 was not very sensitive. The discrepancy between the outcome of the OAT and SI for ADG100, ProtLip, and AMPar illustrates the importance of accounting for genetic correlations between traits in pigs.

4.3 Methodological challenges

Due to necessity, many LCA studies implement a number of simplifications to reduce the amount of data needed to build a model. In livestock LCA models, a system is often described by static inputs and outputs, with feed intake based on feed conversion ratio and animal performance being independent of feed intake, e.g. Kebreab et al. (2016). In this study, we needed to investigate the effects of genetic changes in traits, and feed intake was therefore a function of performance and energy use. Further, we needed to integrate the sow system with the production pig system, as many of the traits frequently targeted by the pig breeding sector have major effects on both the production pig and the sow. To accurately model the effect of longevity traits of the sow, we needed to adapt the model to follow the sow from weaning to culling. This is different from the commonly applied methodology, where the reproductive system is modelled as an annual steady-state system and replacement rate accounts for the number of gilts imported into the system (Nguyen et al. 2011; Mackenzie et al. 2015).

A common argument for not including correlations in LCA modelling is the lack of data availability and the assumption that correlations have minor influence on the outcome (Heijungs 2010; Leinonen et al. 2016). We have shown in this paper that while published genetic correlations for pig traits may be limited in specific areas (i.e. we found no correlations between the trait BWMaturity and traits such as LitterSize, PostWMort and PreWMort), it was still possible through careful weighting to develop a correlation matrix. Instead of changing one variable at a time and accounting for correlations in each scenario (Reckmann and Krieter 2015), we utilized an unguided clustering technique, based on the pooled correlation matrix, to ensure the COT was not affected by personal bias. The limited availability of genetic variances and correlations in literature represented a significant challenge for this study. Conception rates, digestive efficiency, and constants related to the energy required for maintenance by the pigs were all input parameters to the model, but data on genetic correlations and variance of these traits were not available in literature. Major differences have, however, been suggested among pig population for digestive efficiency (Noblet et al. 2013) and heat production (Galassi et al. 2015; Kiarie et al. 2015) which most likely is in turn heavily reliant on maintenance. More studies with non-traditional pig trait genetic correlations, bridging sow longevity and performance, litter size and piglet performance, and grower/finisher performance, would be beneficial for future research on the effects of genetic changes to traits on environmental impacts.

Further, our work illustrated the very different requirements regarding sample size when computing the model uncertainty and the SI. Only 100 samples were required for sufficient precision in the uncertainty analysis, while the SI sensitivity analysis required 20,000 samples to achieve sufficient precision, necessitating multiple hours of computing and fostered a number of coding challenges related to reaching the limits of available computer power. For now, this still represents a barrier for the widespread adoption of global sensitivity analysis in complex LCA modelling as would be considered best practice for quantitative modelling (Saltelli et al. 2008).

4.4 Sensitivity analyses in LCA

The three sensitivity analyses implemented in this study illustrate different applications of the model. The OAT methodology highlighted the importance of every single trait independent of the other traits. Even though OAT is flawed with regard to correlations, the results of this analysis can show the size of the effect each single trait has on the environmental impact of the production system. This has the potential to be utilized to integrate environmental impact considerations and even objectives into breeding programs. The possibilities for using LCA modelling in this way have already been demonstrated for pig systems in relation to diet formulation (Mackenzie et al. 2016; Garcia-Launay et al. 2018).

The COT methodology reflects the genetic correlations among the traits better than OAT and still shows the possibilities of reducing EI in a more realistic way. However, COT is not capable of taking the size of the correlations into account, but can only differentiate between positive or negative correlations. The COT approach can effectively provide a sense check on any strategies devised for reducing environmental impacts, based on changing individual traits from the OAT method. Finally, the SI methodology highlights the capabilities of the traits to affect the environmental impacts, both when looking at individual traits, the interactions of one trait with another and the total influence the traits has on the model. One drawback of the SI method is that the output is an index, which is difficult to translate into an environmental impact reduction and does not show which direction a trait should be selected for to reduce impacts. However, the SI approach can clearly be used to identify the most important areas of focus when devising breeding strategies to reduce environmental impacts, without the need to fully integrate environmental objectives in the breeding model. Even though the SI analysis has been discussed before within modelling and engineering publications (e.g. Abbas and Morgenthal 2016; Kivekäs et al. 2018), and applied to a few animal systems (Godinot et al. 2014; Wolf et al. 2017), it has never been applied to animal LCA modelling. In this study, this method was an essential part of analysing the effect of the genetic correlations among traits.

Multiple global sensitivity analyses are available when conducting LCA models. The methods covered in this paper and possible supplementary sensitivity analyses are shown in Fig. 2. We optioned against using the Sobol index, and the expanded correlated version proposed by Jacques et al. (2006), since the first does not account for correlations and the second assumes there is no correlation between clusters, which in our case was not true. The SI sensitivity method was therefore chosen to account for correlations to their full extend. Since our study did not have access to correlations between the impacts in the primary production system of the feed ingredients in the B matrix, it was not possible to test the sensitivity of this part of the model.

Our study has illustrated the importance of an effective sensitivity analysis that fully encompasses the correlations among the stochastic variables. This is a central point for future studies, since most LCA studies still present limited data on the uncertainties of their models and few have performed sensitivity analysis (Hellweg and Canals 2014). This could partly be due to the limitations of typical modelling programs, which leaves little control of the modelling process and sensitivity test to the user (Mackenzie et al. 2016; Groen and Heijungs 2017; Tallentire et al. 2018). This study proposed a more flexible approach where exports from currently available LCA software were used as inputs to a model on matrix form in coding based modelling software.

The methodology illustrated in this paper is directly applicable to LCA models looking to determine the influence of traits and/or genetic changes on the environmental impacts of a given agricultural system. We have illustrated that a correlation matrix can be compiled from sparse data, and thereby weakened the argument of sparsity of available correlation data as a reason for ignoring correlations (Wei et al. 2015; Groen and Heijungs 2017). A more thoroughly and systematic testing of the sensitivity of biological models, as was done in this paper, could reduce the gap between simulations and reality if implemented into other systems. Further, the implementation of energy animal modelling, together with a similar protein estimation, could be used to predict realistic and optimal diets for more general animal systems than the typical diets often seen in LCA studies.