2.1 Modelling approach and data sources
A mathematical model was written in the Python programming language [27]. This model is a shell program that couples the Livestock Simulation model (LivSim) [28] ,an algorithm to calculate the land footprint of the dairy sector, and a greenhouse gas quantification protocol based on principles of life cycle assessment (Figure 1). LivSim simulates individual cohorts of dairy animals (cows, bulls, juvenile males, heifers, calves) across their lifetime, and the production and environmental impacts (GHG emissions) are aggregated to the production system level. These form the basis for defining a baseline of milk production, emissions, and land use, and for assessing the impact of feeding efficiency gains.
The land footprint indicator includes all land directly used for providing feed biomass: cultivated and grazing land, and land use ‘upstream’ from the farm for production of concentrate feeds. This framework allows an assessment of the impact of changes in diets, or in productivity gains through higher crop yields, to the changes in land use and milk productivity. The dairy land footprint, expressed as hectares per tropical livestock unit (250 kg liveweight), is as forth defined as all crop and grassland directly used for feeding dairy cattle:
![](https://myfiles.space/user_files/58894_9946feeafa4c1df7/58894_custom_files/img1607506730.JPG)
Where b represents the cattle breeds, s represents the livestock production systems, C represents the cattle cohorts, F represents the feeds included in the model, Feed on offer is the annual feed provision per TLU for a given breed, cohort and for a specific feed (Mg TLU-1 yr-1), Yield the annual yield of the given feed (Mg ha-1 yr-1), and Use efficiency the fraction of biomass that is either harvested or grazed. Feed on offer includes all feed available from grazing, harvested on-farm, or purchased from the market.
The model was parameterized with data from a survey of 1,199 smallholder dairy farms conducted in southern Tanzania from November 2017 to August 2018. Surveying activities, performed as part of the IFAD-funded Greening Livestock project, were informed by a stratified random sampling protocol, capturing diversity in dairy farming households (by cattle breed, and socioeconomic factors) among mid to high potential systems across four sampled districts (Figure 2). Baseline indicators characterizing existing feeding practices were developed, which in turn represent diets within the livestock simulations. For the remainder of this paper this survey dataset will be referred to as GLS (2019) [29] .
2.2 Livestock systems and milk production in south and eastern Tanzania
This study focussed on mixed (M) crop-livestock production, rainfed (R), tropical (T) humid (H) systems (hereafter MRT, MRH), following the Robinson et al. (2011) [30] classification. MRT and MRH systems comprise a total of 43,400 km2 (18,500 km2 MRT; 24,900 km2 MRH) across the four regions. In these regions, rainfall is unimodal; the rainy season stretches from November to April, followed by a six-month dry period [31]. Feed sources within these systems depend, to varying degrees, on biomass consumed from grazing, crop residues, cultivated forages, and concentrates acquired off farm. Seasonal variation in feed quantity and quality leads to different grazing and feeding practices across seasons. During the dry season residues from crops form a larger percentage of diets due to the lower availability of natural and planted forages. Concentrates are available from the market year-round but they are generally used sparingly to improve productivity of cows and to maintain nutrient availability during periods of feed scarcity [32]. Protein-dense concentrates, especially sunflower cake, are used to improve milk yields of cows, while maize bran is commonly used as a supplement to maintain energy availability throughout the year [22]. Both of these feeds are produced and processed locally [33] [22]. The baseline diets in the present study, including the seasonal biomass intake from cut-and-carry feeding systems, market purchases, and grazing, were specified using GLS (2019) data (described in SI 1).
The land footprint was disaggregated based on the dominant sources of feed biomass, and the corresponding land uses (Table 1). This allows the impact of changes in croplands and grasslands to land use change emissions to be linked, as per the IPCC (2006) Guidelines. The main feed categories used were: primary crop products (sunflower cake and maize bran), secondary crop products (maize stover), and grass. Grasslands are further divided into native (unmanaged) and sown (managed). The nutritional value and biomass yields of native grasslands were based on the literature on predominant native grass species in the region. Two types of grasses were distinguished based on their yields and nutrient contents: low quality, species of grasses were referred to as ‘Pasture’, which are either harvested or grazed, while ‘Napier grass’ (Pennisetum purpureum), which is the most common improved forage produced in the region (GLS 2019), is considered a high quality, high yielding forage used primarily in cut-and-carry systems.
The fraction of feed available from the total biomass yield, which takes into account the use efficiency, harvesting and manufacturing ratios (e.g. the ratio of bran or cake obtained from the grain or seed portion of the crop) are shown in Table 1. The biomass available from crop residues was calculated using a harvest index of 0.35 [34] .For concentrates the ratio of processed feed products (bran from maize or cake from sunflower) were obtained from literature [35] [36] .The use efficiency ranges from 0.50 to 0.95, and were set to 0.50 for grass and pasture, consistent with values that have been used in previous assessments such as [37 . These values reflect the high stocking rates among highland grazing systems in Tanzania [38] , which result in 0.39-0.61 forage use efficiency [39] .The feed biomass yields per feed type, land use classifications, baseline soil N2O fluxes (see SI for how these were estimated) and C densities of these land use types are shown in Table 1.
Dairy cattle populations and milk production
The dairy sector included all milking cows, replacement females (heifers and female calves), and reproductive cohorts (bulls, juvenile males, and male calves) which are required for maintaining the stock of cows. Between 90-98% of the cows milked in the study areas were indigenous (Bos indicus) cattle, while the other 2-10% were crossbred (Bos indicus x Bos taurus) or purebred (Bos taurus) [40] [41]. Studies indicate that milk production by improved dairy cattle breeds ranges from 1,350-2,200 litres lactation-1 [42] [40] and calving intervals range from 400-520 days [43] [42] .For indigenous cattle, milk yields are typically 500-600 litres lactation-1, and calving intervals range from 450-600 days [42] [40] . Due to the difference in productivity between local indigenous and improved cattle, this study disaggregated the dairy sector (and the dairy land footprint) by breed, resulting in two sectors: the Traditional (local cattle) and Modern (improved cattle) sectors.
2.3 Quantification of greenhouse gas emissions
The dairy sector’s GHG emissions were calculated using an attributional life cycle assessment (LCA) [44] .All major emissions sources were accounted for and calculated in relation to a functional production unit, defined as one kilogram of fat and protein corrected milk (FPCM). Emissions sources included enteric fermentation (CH4), manure (CH4 and N2O), organic and inorganic N inputs into crop and grassland soils (N2O), energy use from manufacturing and transport of feed and fertilizer inputs (CO2), and crop and grasslands expansion (CO2). An allocation factor was used to allocate the emissions from the dairy herd to production of milk and meat, and this value ranged from 0.85 to 0.95. Meat production was calculated using culling rates for each sex (7.7 and 14.0% for female and male cattle, respectively) and a dressing percentage of 52% [40] [45] . Methane and nitrous oxide were converted to CO2 equivalents using global warming potentials of 28 kg CO2eq kg-1 of CH4 and 265 kg CO2eq kg-1 of N2O [46]. Post-farm gate emissions such as for transporting and processing milk were not considered. Thus, the emissions reported are for raw milk produced and consumed at home or sold at the farm gate. Other sources of GHG emissions omitted include those from cattle respiration, farm machinery, electricity, inputs other than feeds and fertilizers, and the construction of farm structures, as these are generally considered minor especially in a low-income context [47] . The procedures followed and emission factors used in calculation of direct GHG emissions are described in detail in the SI. N2O fluxes from crop and grassland soils were modelled based on the IPCC (2006) Guidelines. The results of the baseline values are shown in Table 1.
Carbon dioxide emissions from land use change
Land use changes attributed to changes in feed demand were categorized into one of two transitions: 1) Cropland expansion: grasslands being converted to croplands, and 2) Grassland expansion: other native ecosystems being converted to grasslands. Native ecosystems in this context included wetlands, shrubland, and forests. The CO2 emissions from each transition were estimated using the stock change method [48] [25] . Under this framework, the flux of C (Mg C ha1 yr-1) resulting from the conversion of land is related to the difference in C densities between the current and the previous land use. The C densities for a given land use category are equal to the sum of the five following pools: soils, below and above ground biomass, coarse woody debris, and litter [25] . Following the practice of LUC accounting in dairy LCA, the CO2 emissions after land use change were amortized over a twenty-year period [45] [49]. The transition coefficient for cropland expansion was based on the differences between grassland and cropland C stocks reported in Table 1. This resulted in a difference of 11.0 ± 2.0 Mg C ha-1 between crop and grasslands.
Estimating CO2 emissions from conversion of native ecosystems to grasslands
The extent of grassland expansion was calculated based on the relative availability and utilization of grassland for both LPS based on the density of dairy cattle and availability of grassland per grid cell (see SI for details), following an approach similar to that of Havlik et al. (2014) [50]. Thus, native ecosystems were converted to grasslands when the demand for grasslands exceeds availability. To calculate the transition coefficient, native ecosystem C stocks were estimated using spatially-explicit land cover data at a 100x100m pixel resolution [51] . The C stock density of native ecosystems was estimated as a weighted mean of the shrub, forest, and wetland categories. The C densities of these land categories (for the non-soil C pools) were based on national carbon stock inventory data [52] and for soils, based on a topsoil dataset compiled from 1,400 locations across Tanzania [53] (Table 1). The weights were based on the proportion of shrub, forest, and wetland in a given grid cell [51]. This data was up-scaled to the same spatial resolution as the LPS data and then aggregated to derive a C stock difference between grasslands and native ecosystems representative of both MRT and MRH systems in the study region. The resulting values were 31.5±6.3 and 30.9±6.2 Mg C ha-1 for MRT and MRH systems, respectively. These values are in agreement with the estimates provided by Carter et al. (2018) [54]. LUC emissions from grassland and cropland expansion at LPS level were calculated based on the total amount of land undergoing the given transition in any one year, and the amount of CO2 emitted, after amortization, per unit of land for that LUC transition.
Table 1: Biomass productivity, nitrous oxide fluxes, and carbon density parameters for feed and land use categories in model
Land Use
|
Feed
|
Annual yield
|
Available feed biomass
|
Use
efficiency
|
Nitrous oxide flux
|
Carbon density
|
Mg DM ha-1
|
Mg DM ha-1 yr-1
|
Fraction
|
kg N2O ha-1 yr-1
|
Mg C ha-1
|
Soilsb
|
Other poolse
|
Total
|
Croplands
|
Maize
|
1.46 d,1
|
0.44 (bran)
2.18 (stover)
|
0.95
|
0.73 (stover)
1.03 (bran)
|
38.0
|
3.5
|
41.5
|
Sunflower
|
1.03 d
|
0.36 (cake)
|
0.95
|
0.90
|
Grasslands
|
Napier
grass
|
13.04 a
|
13.10
|
0.75
|
0.51
|
48.0
|
4.5
|
52.5
|
Pastures
|
10.00c
|
3.04
|
0.50
|
0.08
|
Grasslands
|
3.00c
|
1.50
|
0.50
|
0.13
|
Wetlands
|
|
42.0
|
4.4
|
46.4
|
Shrubland
|
|
41.0
|
16.6
|
57.6
|
Forest
|
|
69.0
|
37.8
|
106.8
|
Sources: a[72], b [53], c [8],d [33] ,e[52]
2.4 Scenarios
This study explored three scenarios of improved feeding practices with and without feed crop yield improvements suitable to the agroecological conditions of southern and eastern Tanzania and for each dairy population (indigenous and improved). Similar scenarios were tested previously for Kenya by Brandt et al. [55] [56].This study modifies the scenarios to the policy context and priorities and to the best practice recommendations for the dairy sector in Tanzania (Table 2).
Under the strategy ‘Conservation’ (Cn), urea-molasses treated maize stover was fed to cows in place of untreated maize stover. A urea-molasses treatment is proposed to enhance the nutritional quality of stovers [11]. Therefore, in the dry season when availability and nutrient quality of forages is reduced, feeding treated maize stover can increase protein intake. The ‘Forage’ strategy (Fo) evaluated the role of higher rations of Napier feeding, in place of grass and pasture. For the ‘Concentrate’ strategy (Co), supplemental concentrates were provided to cattle according to supplementing regimes aimed at optimizing milk yields for local and improved cattle [57] [28]. All three of these strategies were evaluated additively by first implementing the conservation strategy, then assessing the additional effect of Fo and Co. This is because feeding greater concentrates was not found to be effective in improving milk yields unless seasonal feed deficits were first reduced (e.g. by using feed conservation and greater forage quality). For the results of additional scenarios, and the seasonal variation in nutrient availabilities for the cow simulations, see SI Section 6.
The Tanzanian Grazing-Land and Animal Feed Resources Act [58] seeks to catalyse the development of Tanzania’s commercial feed processing industry. The simulations therefore focussed on yield gains in maize and sunflower for concentrate production, which are the two most common sources of concentrate feeds in the region [22]. Current yields of these crops (Table 1) are significantly below their potential, with water limited yield potential having been reported up to as high as 6.0 (maize) and 3.0 (sunflower) Mg ha-1 yr-1 [59] [60]. Data from field experiments in Western Kenya [61] were used to estimate the effect of higher N fertilizer application on yields and N2O emissions of maize and sunflower used in concentrate production. The yield gains were set as 50% of the yield gap based on the values reported above and in Table 1. The fertilizer requirement used to achieve these yields were based on an N-yield response of 14 kg ha-1 kg N-1, with an emission factor of 0.015 kg N2O kg N-1 [61]. These scenarios were implemented in addition to the above feeding strategies, and denoted with a ‘+Cyg’ (‘Crop yield gains’). The results of the yield gap and N2O calculations used for these simulations are shown in SI 4.
Table 2: Definitions of scenarios examined and their target populations of cattle
Sector
|
Cattle
population
|
Feeding strategy
|
Scenario abbreviation
|
Description
|
Traditional
|
Indigenous
|
Conservation
|
L - Cn
|
All maize stover fed to cows is treated with urea-molasses.
|
Conservation plus forage quality
|
L - CnFo
|
L – Cn with Napier grass increased to 25% of feed on offer, replacing grass and pasture.
|
Conservation plus forage quality with supplementation
|
L - CnFoCo
|
L - CnFo with 2 kg of concentrates fed during early lactation, and 0.5 kg per day during other periods. Concentrate intake is comprised of 67% maize bran and 33% sunflower cake.
|
Modern
|
Improved
|
Conservation
|
I - Cn
|
All maize stover fed to cows is treated with urea-molasses.
|
Conservation plus forage quality
|
I - CnFo
|
I – Cn with Napier grass increased to 50% of feed on offer, replacing grass and pasture.
|
Conservation plus forage quality with supplementation
|
I - CnFoCo
|
I - CnFo with supplement feeding involving
5.0 kg d-1 of concentrates during early lactation, and 1.5 kg d-1 during other periods.
Concentrate intake is comprised of 67% maize bran and 33% sunflower cake.
|
Baseline production growth and greenhouse emissions
A baseline provides a reference level against which a mitigation goal can be established [9]. The production practices used in the baseline represent those in the absence of specific mitigation interventions [62].The dairy herd population for 2020 was establishing using spatially-explicit data on livestock population densities [63] and annual growth rates in herd size. Feeding practices were obtained from GLS (2019) (SI 1). Model parameters for the Baseline were thus set by extrapolating historical values over the 10-year timeframe of the assessment. Throughout the 10-year simulation period, the herd size was assumed to grow by 5.5% and 4.5% annually for local and improved cattle, respectively [64]. No changes were assumed for feeding or other herd management practices that would otherwise affect productivity or herd compositions. The yields of feed crops were assumed to grow consistently with historical averages of 3.4% and 4.1% annually for maize and sunflower, respectively [33]. The scenarios were run modifying the availability of feeds, with and without yield improvements. For these scenarios, the populations and herd structures remained constant. The scenarios described above for both Traditional and Modern systems were thus run to compare to the Baseline scenario. This resulted in a total of 14 runs (2 baselines + 2 sectors x 3 feeding scenarios x 2 crop yield variants) for each LPS.
2.5 Uncertainty assessment
Uncertainty in GHG emissions was quantified in line with the IPCC (2006) Guidelines. In the baseline, the sources of uncertainty were dairy cattle numbers per LPS, feed on offer per head, biomass yields, and emission factors (including coefficients on LUC transitions). For subsequent simulations the dairy herd and feed intakes were specified in relation to the baseline, and therefore for all other scenarios the only sources of uncertainty were in emission factors and biomass yields. Monte Carlo (MC) simulations were run for the baseline and each subsequent scenario to estimate the GHG emissions error range at a confidence interval of 95%. The standard error in emission factors were specified based on IPCC (2006) Guidelines. The uncertainty in enteric fermentation, which was calculated using Tier 3 guidelines, was set at 10%. The coefficients for LUC were calculated from country specific inventory studies and thus were either Tier 2 or 3 emission factors [52] [53]. Moreover, because these coefficients were highly dependent on the C density data reported by Mauya et al. [52], who report relatively low uncertainty (0.9% for forest and 1.8% for non-forest land), the standard errors for such were set at 20%. Because this study included simulations for greater N-fertilizer application, which may result in highly variable and uncertain changes in N2O emissions, the standard error of this emission factor (EF1 soil N inputs) was set at ± 66%. All other emission factors ranging from Tier 1 to 3 were set based on IPCC guidelines, thus ranging from 7 to 30% (SI 5).