Biophysical Simulation of Sheep Grazing Systems Using the SGS Pasture Model

Abstract

The performance of farming systems models for grazed grasslands are seldom evaluated against comprehensive field data. The aim of this study was to evaluate the capacity of a daily time step, grazing systems simulation model—the SGS (Sustainable Grazing Systems) Pasture Model—to simulate production and aspects of sustainability. This was completed by evaluating temporal changes in soil water balance, some major nitrogen (N) fluxes, as well as plant and animal production using data from two large scale experimental sites with grazing sheep. The simulations were broadly in agreement with the measurements. In cases where divergence occurred the reasons were apparent and could be explained by reference to the model structure or aspects of the field data. In particular, the simulations showed good agreement with the observed soil water, but poorer agreement with the volumes of runoff. The simulated N in leachate and soil inorganic N were less in agreement with the measured data. The model outputs were sensitive to symbiotic biological fixation by subterranean (sub) clover and mineralisation of soil organic matter, which were not measured. Similarly, there were variable results for the simulation of animal growth and production. The complexities of simulating grazing systems and comparing field observations to simulated values are discussed.

1. Introduction

Grazing systems are complex agroecosystems, and the variable responses of many of their components to climate, soils and management means they can often be difficult to comprehensively interpret or predict using experimental methods. As a result, the popularity of biophysical simulation models for understanding and forecasting grazing systems has increased markedly in recent decades [1]. The use of grazing systems models operating at the paddock to whole farm level is now a key step in assessing the effects of changes, for example, intensification, climate change, drought, tactical management, mitigation of environmental pollution, on actual and hypothetical systems [2,3].

A range of models are currently used to comprehensively represent grazing systems. In Australia and New Zealand, these include the SGS Pasture Model [4], DairyMod, [5], GRAZPLAN [6], GrassGro [7], DairyNZ whole-farm model [8] and further abroad, other models [9,10] with similar objectives. The component models forming the foundation of these grazing systems models are often complex representations of specific processes such as the flow of nutrients, energy or water throughout plants, animals and soils. When these are combined, the complexity of a grazing system model is often significant due to the degree of the mechanistic representation of the processes being simulated, the number of processes being simulated, and sometimes, due to the number and availability of inputs needed to execute the model. The complexity increases further as the models change spatial scales from point scale to sub-paddock and to farm-scale.

The mechanistic component models of agroecosystems models are often developed, calibrated and even comprehensively validated using data gathered from closed experimental conditions such as laboratories, glass or animal houses. As more component models are linked together and simplified, however, testing and evaluation becomes increasingly more difficult. Consequently, the outputs from large, interconnected agroecosystem models are rarely comprehensively validated, owing not only to the complexity of the task but also due to the fact that the necessary experimental datasets are rarely available. Simulations, as a result, are usually interpreted based on sensibility or plausibility [11] with reference to the reasonable bounds of what might be expected by experienced researchers e.g., [12]. Even though a number of assessments of the outcomes of farming systems models have been made against experimental data, these are usually only for selected parts of the models e.g., [13,14,15,16,17]. Often the performance of a selected part of these complex models (e.g., plant production) is inferred to indicate the adequacy of the model to represent the whole grazing system e.g., [18].

There is an ongoing challenge in how best to use data from grazing system experiments to validate models. There can be a disconnect between what is sampled or measured and how it is interpreted when it comes to modelling [19]. One of the challenges in the use of field data to assess the performance of models is that biophysical data often exhibit a large degree of variability. Discrepancies between predictions and observed data caused by natural variation are also common. Any model validation must therefore be accompanied by a critical examination of the source and nature of the accuracy and quality of the data used [20]. Data are often collected in diverse ways at varying intervals and spatial scales than those represented in the models which can be challenging when using models and data together e.g., [16]. Most often data are not collected with model testing in mind and so modellers often must “work with what they’ve got”. Data from field research may also have an uncertainty associated with experimental design and error, limitations in measurement techniques and incomplete information such as the limitations of locally important factors on soil fertility [14,18,21,22]. Some of these challenges include inadequate descriptions of one or more key factors required for model validation, differences in data collection methodologies or frequencies, instrument calibration, etc. Often, the information is reported as the mean of the data across treatments or stocking rates, or dates or years. Rarely are precise pasture and animal data reported from the same experiment [16]. Albeit problematic, the evaluation of simulations together with data is an essential step in the modelling process.

Other than the dataset from the Sustainable Grazing System (SGS) project [23], an extremely limited number of detailed datasets exist of complex grazing systems experiments (i.e., covering plants, animals, water, soil). The purpose of this paper therefore is to evaluate the performance of a grazing systems model against measurements of soil water balance, soil N, plant and animal production in different sheep grazing systems in southeast Australia. In doing so, the principles of what is required of models to represent experimental data from grazing systems are illustrated by exploring examples of the limitations of both modelled and experimental data.

2. Materials and Methods

The dataset used in this evaluation exercise came from the SGS experiment. The SGS experiment conducted from 1997 to 2001 broadly aimed to investigate various aspects of productivity and sustainability of grazing systems in southern Australia [23]. The SGS Pasture Model is a grazing systems model [4] that was developed to enable the simulation of complex interactions between the climate, soil, nutrients, water, pasture species and grazing animal management. It is described in detail in [4] and http://imj.com.au/sgs/ (accessed on 24 November 2022) and will not be elaborated on further here.

2.1. Sites and Data Simulated

Experimental data recorded from the SGS experimental sites located at Ruffy (36°37′ S, 145°70′ E) and Maindample (36°59′ S, 145°59′ E) in southeast Australia from 1997 to 2001 were used in the model evaluation. The two locations are about 50 km apart and of similar elevation above sea level and differ according to soil type, pasture base and breed of sheep. As the sites located within 50 km of each other they share a similar climate with a 5% median difference in average annual daily temperature and a 3% median difference in annual rainfall. Both sites were managed as commercial grazing properties prior to and during the study. The design of the field experiment, fertiliser history and methods of measurement for the various system components are detailed in [24,25,26]. The layout of the experiment is shown in Figure 1.

The experiment compared a range of systems based on soil P fertility: Low, Medium and High. (Note that throughout the paper the fertiliser treatments at each site are abbreviated as L, M and H, respectively. These are preceded by the first letter of the site such that MH is the High fertiliser treatment at Maindample, and RL is the Low treatment at Ruffy. Fertility treatments are indicated by a capitalised word). The Medium and High systems came from a division of one large paddock. The monitored area (called catchments in Figure 1) were 4.9–13.7 hectares at Maindample, and smaller at Ruffy from 1.8 to 3.9 hectares. The Medium and High treatments at each site were original paired-paddocks and both were sown pastures. These treatments were compared with Low (control) an area where the pasture had not been sown and had typically received little fertiliser in the twenty-five preceding years.

Each site was located within an existing paddock. Within each paddock, a natural sub-catchment (hereafter called a catchment) was surveyed and then hydrologically isolated so that water and nutrients could be collected from surface runoff and subsurface lateral flow.

The Low system came from a separate paddock that had a different land management history to the other two systems. The grazing systems were based on the variable annual applications of superphosphate to the grazed area. The systems were replicated at both the locations. At Maindample they were 5.5 kg P/ha for Low, 11 kg P/ha for Medium and 22 kg P/ha for the High system. At Ruffy they were 4.5 kg P/ha for Low, 9 kg P/ha for Medium and 25 kg P/ha for the High system. The Medium/High systems at both sites were limed in 1992.

At Maindample the soil has been classified as a Brown Sodosol [27]. It has a strong texture contrast between the topsoil and subsoil, with a bleached A2 horizon and impermeable sodic subsoil. At Ruffy the soil has been described as a bleached-mottled magnesic Yellow Kurosol [27]. Both are typical of soils used for sheep farming in many parts of southeast Australia. Further details of the various soil properties at the beginning of the experiment are given in [24].

The MH and MM systems were sown to the improved pastures of Phalaris cv. Australian (Phalaris aquatica L.) in 1972 and 1975, respectively. A prime lamb enterprise was simulated consisting of a Dorset ram over a Border Leicester–Merino first cross ewe. Ewes were joined for 6 weeks from mid-January and lambed in June. In December, lambs were weaned for sale and ewes were shorn. The RM and RH systems were sown to Cocksfoot cv. Porto (Dactylis glomerata L.) and subterranean clover (Trifolium subterraneum L. cvv. Trikkala and Larisa) in 1989. A self-replacing Merino ewe mob was run. In December, lambs were weaned and ewes were shorn. The Low systems at both sites were based on unimproved native annual and perennial species and were un-limed.

Data were collected at time steps that varied from sub-daily to annual. Depending on the measurement these differed and are detailed in

Table 1. Summary of experimental data used in the evaluation of the SGS Model (further details such as methods of measurement are shown in [24].

Data	Frequency of Data Collection
Climate
Rainfall	Recorded in 0.2 mm tips using one automated weather station at each site
Weather data	SILO data drill (http://www.longpaddock.qld.gov.au/silo/ accessed 25 June 2022) and FAO56 evapotranspiration
Soils
Saturated hydraulic conductivity (field and laboratory methods)	Recorded using disc permeameters once from field pits at three locations in each system
Saturated moisture content and field capacity	Recorded once on intact samples from field pits at three locations in each system (NB there was no permanent wilting point and air-dry moisture data recorded)
Pasture
Dry matter (total and for major plant groupings)	Above ground biomass (live and dead) recorded at each of the measuring stations on a seasonal basis (approximately 4 times a year)
Botanical composition analysis	(As above)
Metabolisable energy, crude protein, digestibility	(As above)
Live and dead material	(As above)
Animal
Mature animal weight	Recorded on a seasonal basis (approximately 4 times a year)
Fleece weight	Annual mean for the mob at shearing time
Lamb weight	Recorded annually at removal/sale
Lamb number	Recorded on a seasonal basis
Water
Runoff and lateral flow volumes	Recorded in flumes at the bottom of the catchment on a runoff-event basis
Soil moisture content	Recorded on a weekly basis at regular intervals (20, 40, 60, 80, 100, 120 and 140 cm below the surface) at each measuring station
Nutrients
Soil inorganic N concentrations	Recorded on a weekly basis at regular intervals (20, 40, 60, 80, 100, 120 and 140 cm below the surface) at each measuring station
Nutrient concentration (nitrate, ammonium in suction cups at 140 cm)	2–3-week intervals at each measurement station

.

Further details of data collection methodologies can be found in [24]. Surface runoff was collected at the lowest point in the landscape of each system. A barrier was inserted into the soil across the bottom boundary of each catchment to channel surface runoff into a flume. Subsurface lateral flow was collected by a trench that was dug to the depth of the subsoil. More details of the runoff collection and partitioning are given in [24]. Separate flocks of sheep grazed each system continuously at both sites, except for a 6-week period when system flocks were grazed together when the ram was present in the mob. Stocking rate was set at 1 DSE/ha for each 25 mm above a baseline level of 250 mm and then adjusted according to the fertiliser application. Stocking rates from 1998 to 2000 are shown in

Table 2. Mean stocking (ewes/ha) and lambing rates (lambs/ewe) for different grazing systems from 1998 to 2000. (Note: The mean values for each system used for simulations).

Year	RL	RM	RH	ML	MM	MH
	Ewes
1998	5.9	7.3	10.9	4.2	6.0	7.6
1999	5.9	7.0	10.9	4.9	6.4	7.7
2000	6.2	7.2	10.9	4.9	6.4	7.6
Mean	5.9	7.2	10.9	4.6	6.3	7.6
	Lambs
1998	0.9	0.8	0.8	1.43	1.39	1.36
1999	1.1	0.7	0.8	1.40	1.43	1.38
2000	0.8	0.5	0.7	1.29	1.41	1.36
Mean	0.9	0.7	0.8	1.37	1.41	1.37

. In the SGS field experiment these stocking rates were set with the objective to produce a similar weight gain/animal in each system [24].

2.2. Whole System Simulations and Parameter Value Estimation

The SGS Model v. 4.8.16 [4] was used to simulate the grazing systems. The model was not developed based on the experimental outcomes and the default parameter sets largely came from the literature; therefore, the comparisons between the model and experimental data were independent. The model includes modules for pasture growth, pasture utilisation by grazing animals, water and N dynamics, animal and plant physiology and production. There are a range of options to realistically represent farm activities such as pasture management, feeding and fertiliser application. Each paddock can have up to five pasture species, which can be annual or perennial grasses, C3 or C4, or legumes or forbs. The model is complex and so rather than provide a detailed description here, the reader is referred to further descriptions and documentation of the model [4,28].

Data collected between the years 1997 and 2001 were maintained in the SGS database [29] and were used here to set-up and evaluate the model. Two sheep-based enterprises were simulated: a prime lamb system at Maindample and a wool production system at Ruffy. Details of the mean stocking and lambing rates are given in Table 2. A generic pasture sward was simulated based on Phalaris and subterranean clover even at Ruffy where the sward was cocksfoot. Although the actual number of species present was much greater [24,30], the approach of a functional group and a simple approach were adopted with the three species representing all plant species in the whole sward. Despite subterranean clover often being a minor component of the total pasture dry matter (median of 5% at Ruffy and 11% at Maindample), it was a necessary inclusion in the simulation to provide an input of N (via symbiotic fixation) to the systems. Otherwise, there would not have been an external source of N to sustain the plants. To allow soil carbon and N within the model to reach a steady state, the simulations were run from 1972.

Historical daily climate data were obtained from the SILO database (http://www.longpaddock.qld.gov.au/silo/ accessed 25 June 2022) for the years 1972 to 2000. The climate data inputs included minimum and maximum temperature (°C), rainfall (mm), solar radiation (MJ/m2) and vapour pressure (kPa) and potential evapotranspiration calculated using the Penman–Monteith (FAO56) method.

Where available, data from the field experiments were used to set up the model and to parameterise each simulation for the different systems. The parameter “daily rainfall distribution” in the model refers to the number of hours during which precipitation falls in a day. The daily distribution was set equal to the duration of the mean rainfall event for the experimental period, which was determined to be 4 h. Pregnancy, lactation and growth characteristics differed between sites to reflect differences between the two sheep breeds. The model does not have a detailed representation of the mob structure (i.e., differences in age, reproduction performance) and has fixed lambing rates for all animals in the mob. The milk production of the two breeds was based on the literature values [31,32]. As noted by [22], ideally, the parameter values would be estimated directly from experimental data but when not available it is appropriate to adjust these parameters to gain agreement with the data. However, most authors [14,18,21,22] usually find it necessary to modify a limited number of plant parameters to reflect the species, breeds, cultivars or local conditions. Indeed, the transparency of the model interface has often been noted as one of its strengths [4] but parameter values are seldom modified. Additionally, due to the low lambing rate at Ruffy (less than 1.0 lamb/ewe) some adjustments of the parameters were needed to reflect the overall mob as the minimum.

Other than the parameters used to set up the simulations (e.g., soil properties, plant species, milk production, lambing rates), following some initial model runs, a limited number of key parameter default values requiring alteration were identified. These ensured better agreement between the simulations and the experimental data. This was expected as many of the parameters are generic to increase model utility for users, but for experimental purposes they need to be tailored for local conditions. This was performed in a systematic and iterative way to develop a parameter set rather than individual parameters. In some cases, this was based on published and prior knowledge of these and similar pasture-based grazing systems and other times it was based on trial and error. A protocol was undertaken to ensure that the model was not over-calibrated for a local dataset. Where the actual value of the key parameter default in question was being reviewed, the first year of the experimental data in the High system was used to determine a value to produce a low prediction error. This was then fixed for the remainder of the experimental period and assessed against data from the other systems at that site, and across both locations. Where a widespread improvement across the sites and systems occurred, the value was set. On occasions where certain parameter values (or sets of values) gave a better fit of some parts of the system (e.g., total biomass production) they were checked (and changed as necessary) against other parts of the system (e.g., soil moisture in different layers; animal weight) and against other systems at the same site to ensure the best fit of the data overall across a range of different aspects of the system. In this way, the final parameter set (a combination of parameter values; shown below) gave the best fit of the data overall.

A comprehensive sensitivity analysis was not conducted. However, the effects on predictive capacity due to the alteration of a few key variables (a parameter set) were assessed against the default variables by using the final parameter set compared to the default parameter set. The effects of the parameter changes on the model outputs were illustrated using data from the Medium input system at Maindample.

2.3. Data Analysis

The overall model performance was based on the version with the modified parameter set (i.e., all the parameter changes combined) rather than individual changes. As per other authors [18,21,22], combinations of objective and subjective measures were used to assess the model outputs. Observed field measurements were compared to model simulations graphically and statistically using Genstat version 14 (Lawes Agricultural Trust 2011). Apart from traditional statistical tests, for complex models with many interlinked components, it is also important to examine the general behaviour of the model over time, drift, ability to deal with extremes and spikes [22]. This ensures general consistencies with patterns and the mean behaviour of actual systems such as the timing of variations between simulations and observations to assess trends. Comparisons were made between temporal changes in soil water, soil N, plant and animal production simulated by the model and the data from entire experimental period. Spearman’s rank correlation (coefficient r) was used to test the association between actual values and simulations.

Observed values (Oi) and predicted values (Pi) were also compared using root mean square prediction error (RMSE) for all observations (n) of each variable:

$$RMSE={[\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(p_i-o_i\right)}^2]}^{\frac{1}{2}}$$

This gives an indication of the overall prediction error. The square root of this (RMSE) was used to assess the goodness of the predictions. Significance was attributed at p < 0.05.

The data from the first year were not used in the statistical analysis of the predictions except for the plant measurements and the lamb data as there were few observations of both sets of data.

3. Results

3.1. Parameter Value Estimation

The set of parameters which were modified during this process are listed in

Table 3. Changes to the model defaults for the grass (non-legume) and legume component (sub-clover) defaults in the different grazing systems.

Parameter	Default	RL	RM	RH	ML	MM	MH
	Non-legume component
Maximum root depth (cm)	150 (40)	55	90	95	95	95	95
Depth for 50% root distribution (cm)	30 (15)	30	25	15	15	15	25
Minimum temperature for leaf flux (°C)	0	8	8	8	9	9	9
	Legume component
Legume GLF N at zero available soil N (%)	60	60	60	80	60	80	80
% N fixation at full soil N availability	40	50	40	70	40	60	60
	Animals
Normal mature body weight (kg)	50	66	63	63	77	77	77
Minimum mature body weight (kg)	30	40	40	40	61	61	61
Days from birth to removal	60	112	121	124	166	166	166
Peak milk production (L/d)	1.6	1.23	1.23	1.23	2.10	2.10	2.10
Wool potential growth rate (g/day)	15	17	16	18	15	17	15
Growth characteristic—negative exponential scale factor (days)	150	120	130	150	120	120	120

and explained below; however, many more were assessed. Of the very many model parameters, these are a minimal number.

The extreme temperature stress effects had a profound effect on pasture growth. During 1997–2001, 21% of days had a minimum temperature ≤5 °C, and 13% of days had a maximum temperature ≥30 °C. Implementing the stress meant that leaf photosynthesis capacity and respiration were limited. This is important for example in winter, as this region regularly experiences severe frosts that adversely affect plant growth. This modification meant that plant growth was influenced by extreme temperature for more than one third of the experimental period. Implementing (turning on) the default low temperate effects (see Table 3) resulted in small amount of dry matter production between May and November (Figure 2). Turning off the effects of high temperature led to an over simulation of pasture biomass production from December through to February. Therefore, the high temperature effects were used in subsequent simulations; however, the low temperature effects were not.

Modifying the parameter called “minimum temperature for leaf flux” had a large effect on the amount of plant material produced. Changing this effectively altered the speed with which the plant material passed through age categories. In particular, the under-prediction of plant growth over the winter period was improved, which led to an increase in accumulated dry matter of Phalaris from May to November, and for the clover from July to November. This change ensured that the ratio between live and dead material from end of April to end of November was in better agreement. This change did not appear to have any detectable impact on animal weight which might have been expected.

Changing the N fixation parameter was necessary to simultaneously improve the simulation of legume and grass dry matter production, legume and grass N%, soil inorganic N and N leached. This is valid as although the %N fixed changed within a location, this was expected as the symbiotic fixation is affected by a number of physical (i.e., water logging) and chemical properties of soil (i.e., soil pH, nutrient toxicity/deficiency). Without a replicated study it is not possible to say if this was a site specific/local effect or a response to the farming system (combination of soil fertility, liming, companion species, grazing pressure). The final parameter values provided the best agreement across these various factors. Modifying the N fixation parameters resulted in a 26% increase in growth of Phalaris and a 10% increase in clover over the 5-year period compared to the default (Figure 2) and a 24% increase in the amount of N fixed and a 70% increase in the amount of N leached. There was no effect on animal weight changes as result of modifying the N fixation parameters, even though there was an average increase of about 1% in the protein concentration of the forage and an increase in the biomass.

Changing the default rooting depth was justified based on the soil water data. Changing the maximum rooting depth had a 6 % difference on dry matter production overall for the 5-year period. The greatest effects on plant production were observed over summer and autumn for Phalaris in particular (Figure 2). There was an annual average difference however of 12% (or 45 mm) in pasture transpiration.

Modifying the “dead to litter” resulted in less dead plant material consumed from Feb to May and less overall plant material from December to April. Although this resulted in an overall increase in herbage ME of 1%, there was an insignificant net effect on animal weight changes as one effect cancelled out the other. Modifying the “sheep growth” parameter influenced the efficiency with which resources were used. Consequently, there were no differences in pasture intake but there were differences in animal growth patterns during the period between lambing and joining. Changing the wool production parameters resulted in a 12% difference in fleece (greasy) production on average or 500 kg/ewe. Modifying the milk production parameter had a minor effect on the pasture intake (300 kg difference) and lamb weight (500 g). Any differences in soil properties translated into insignificant differences in plant biomass production but large differences in soil water drainage over the duration of the experiment.

The effects of altering a range of other parameter default values were assessed; however, they had insignificant effects on the simulations compared to the default parameters.

Overall, details of the agreement between observed and simulated data for the model components evaluated using the final simulations compared with data collected from the grazing systems are shown in

Table 4. Overall ranked correlation coefficients for the different grazing systems for 1999–2001. The data for lambs are for the period 1997–2000. (Numbers in brackets are the RMSE).

	RL	RM	RH	ML	MM	MH
	Runoff
Surface 1 (mm)	0.20 (0.6) *	0.22 (0.7) *	0.40 (0.5) *	0.62 (1.1) *	0.26 (1.8) *	0.31 (1.3) *
Surface + Lateral 1 (mm)	0.20 (0.7) *	0.22 (0.8) *	0.40 (0.4) *	0.65 (1.3) *	0.26 (2.0) *	0.31 (1.3) *
Soil Water
20 cm (%)	0.85 (2.5) *	0.83 (2.6) *	0.75 (3.1) *	0.91 (3.2) *	0.74 (5.9) *	0.79 (5.8) *
40 cm (%)	0.81 (3.6) *	0.54 (3.2) *	0.54 (3.5) *	0.66 (4.0) *	0.65 (4.0) *	0.75 (4.1) *
60 cm (%)	0.72 (4.3) *	0.66 (1.6) *	0.72 (2.5) *	0.68 (3.2) *	0.65 (2.5) *	0.78 (2.3) *
80 cm (%)	0.57 (1.1) *	0.65 (0.8) *	0.61 (1.5) *	0.70 (3.8) *	0.62 (1.9) *	0.77 (3.0) *
100 cm (%)	0.43 (2.2) *	0.45 (0.7) *	0.49 (1.8) *	0.57 (3.0) *	0.53 (1.1) *	0.81 (24.9) *
120 cm (%)	0.30 (0.7) #	0.14 (0.8) n	0.09 (0.7) n	0.22(1.1) n	0.09 (0.9) n	0.56 (10.9) *
140 cm (%)	0.29 (0.7) *	0.03 (0.8) n	0.17 (0.6) n	−0.21 (1.6) n	−0.23 (1.1) n	0.27 (2.3) #
			Plants
Total dry matter 2 (t/ha)	0.82 (0.8) *	0.83 (0.6) *	0.79 (1.2) *	0.79 (1.6) *	0.71 (1.1) *	0.62 (1.4) *
Clover dry matter (t/ha)	0.75 (0.8) *	0.85 (0.5) *	0.72 (0.4) *	0.88 (2.1) *	0.85 (0.5) *	0.90 (0.6) *
Live dry matter (t/ha)	0.83 (0.6) *	0.90 (0.5) *	0.78 (1.0) *	0.87 (0.7) #	0.95 (0.6) *	0.91 (0.9) *
Dead dry matter (t/ha)	0.61 (1.3) *	0.85 (0.3) *	0.91 (0.4) *	0.24 (1.4) #	0.48 (1.2) #	0.47 (0.8) #
N (%)	0.75 (0.6)#	0.93 (0.8) *	0.96 (0.5) *	−0.07 (2.2) n	0.85 (0.6) *	0.32 (1.0) n
Digestibility	0.39 (14.0) n	0.43 (9.2)	−0.05 (11.2) n	0.95 (5.1) *	0.95 (6.8) *	0.51 (13.3) n
Metabolisable Energy	0.38 (1.4) n	0.43 (1.4) n	−0.05 (1.4) n	0.95 (1.3) *	0.95 (0.6) *	0.54 (1.6) n
			N in leachate
Nitrate (mg/L)	−0.72(11.4) *	0.11 (7.2) n	0.47 (2.4) *	−0.37 (12.5) n	0.29 (3.0) n	0.20 (7.6) n
Ammonium (mg/L)	−0.14 (0.1) n	−0.49 (0.1) n	−0.01 (0.1) n	0.26 (0.3) n	0.35 (0) n	0.29 (0.1) n
			Animal 3
Ewes (kg/animal)	0.10 (6.4) n	−0.22 (8.2) n	0.22 (6.9) n	0.48 (5.1) *	0.28 (5.4) n	0.47 (5.6) *
Lambs (kg/animal)	0.98 (4.2) *	0.98 (3.4) *	0.91 (3.2) *	0.88 (3.1) *	0.98 (2.3) *	0.92 (2.3) *

¹ Only for days when rainfall was recorded, ² Total standing biomass (live and dead), ³ Daily weight gain, * values are p < 0.05, ^# values are p < 0.1, ⁿ values are p > 0.1.

. The correlation values were high and positive in most cases for water, plant and animal outputs. Further details are provided in the following sections. Different systems have been shown for different themes (e.g., water, pasture, animals, etc.). The purpose of the selections was not to hide any discrepancies but to highlight relevant differences or similarities between the systems.

3.2. Water Balance

The agreement between the observed and simulated soil volumetric data was good for all systems, for all years and for all soil depths. The simulated and observed data had the highest correlations for the 20 cm layer (r ≈ 0.8). In the bottom layer (140 cm), soil moisture was simulated within the range of the observed data, and there was very little variation throughout the experimental period. However, a correlation is not the best measure of the model performance as the correlation was low due to limited range in the soil moisture over the period. For example, in the case of the Ruffy Medium system the standard error of the mean volumetric soil water content over the experimental period at the 140 cm depth was 0.8. In cases where correlation coefficients were low (e.g., below 100 cm depth), there were narrow ranges in the mean soil moisture throughout the year and with a small RMSE. For the time series of the predictions, visually the layer with the weakest agreement was at the 40 cm layer which corresponded to the top of the clay subsoil (e.g., Figure 3). Observations of soils pits at Maindample indicated this was the layer for the movement of water laterally through subsurface pathways. This was not so obvious in the soil profiles at Ruffy, which had a more permeable profile overall [33].

Overall, the period of strongest agreement with the observed soil moisture was from May to September (Figure 3). The greatest differences were found in spring (Sept, Oct, Nov), with the model drying the soil more quickly, and with a lower content of soil moisture, than the mean of the observed data.

For the experimental period there was a large under-prediction of runoff days compared to the observed data at all sites. The ambiguity in the observations concerning the definition of daily runoff was more evident at Maindample than Ruffy. For example, for the Maindample Medium system surface runoff was recorded on 12 days when rainfall did not occur. For the same system, for subsurface and lateral flow combined, the number of days was 30. Consequently, the adjusted runoff was used for evaluations of the model outputs; therefore, only the runoff recorded on days when precipitation was also recorded was used for evaluation.

At Maindample there was large variability between systems with the number of runoff days observed in the Medium system consistently double that in the High system. At Ruffy, Low surface runoff was less than 0.5 mm for around 89% of days, and subsurface lateral flow was less than 0.5 mm for 87% of days. When only the adjusted runoff values were summed, the annual runoff volumed were broadly comparable between observed and predicted data for all sites at Maindample and Ruffy in all years, possibly with the exception of year 2000 (

Table 5. Observed and predicted total annual runoff (mm) and number of surface runoff days for the different grazing systems from 1998 to 2001.

	System	Observed	Predicted
Runoff (mm)	ML	127	103
	MM	220	122
	MH	107	129
	RL	18	6
	RM	13	4
	RH	7	5
Runoff Days	ML	11	13
	MM	50	16
	MH	21	13
	RL	2	1
	RM	4	5
	RH	3	1

).

The prediction of the timing of large runoff events (e.g., >1 mm/d) was mostly consistent with the observed data, but inconsistent for the considerable number of small events (e.g., <0.5 mm/d). This was more obvious at Maindample than Ruffy. For surface runoff, the correlations between adjusted daily runoff predictions were sensitive to the saturated hydraulic conductivity of the surface layers and the number of hours of rainfall per rain-day (parameter “daily rainfall distribution”). The slope value used in the simulations was an average for each of the areas (however, runoff was mostly insensitive to this parameter). The dominance of small runoff events from July to November in 2000 in response to a period of extended rainfall meant that surface runoff also was overall poorly predicted.

3.3. Nitrogen in Soil and Leachate

Soil nitrate and ammonium concentrations throughout the profile measured annually in April each year were predicted within the range of the observed data for most depths on the dates when measurements were taken (Figure 4).

Observed solution nitrate concentrations were highly variable, as expected under grazed pasture [26], and were not well predicted at the Ruffy site. Daily nitrate in the drainage water (when drainage was greater than 0.1 mm/d) at the same depth as the suction cups was derived by dividing the N leached by the drainage volume. Compared to the observed concentrations of nitrate in suction cups, there was a general under-prediction throughout the experimental period (even with the adjustment of the N fixation parameters). The example shown in Figure 5 indicates that the disagreement between observed and predicted data decreased during the latter half of the experimental period. There was also an under-prediction of ammonium leaching at both sites.

3.4. Pastures

Overall, the model predicted the plant production well during the experimental period with all correlation coefficients between observed and predicted being greater than 0.50. There was an over-prediction of total sward dry matter in late spring (e.g., November) in the last two years of the experiment and an under-prediction in early mid-autumn (e.g., Figure 6a). The growth dynamics of key sward components such as the biomass of the functional groups and proportion of grass:non-legume were well replicated by the model at all locations. It is noteworthy that this period of over-prediction coincided with the period when the predicted soil moisture notably deviated from the observed data. At Ruffy the agreement gradually worsened until the time that the RH and RM systems were re-sown in autumn 2000.

The legume (sub-clover) dry matter predictions were in close agreement with the observed data (Figure 6b). The total sward N concentration (above ground tissue) was calculated from the dry matter proportions of the non-legume and legume components and their N respective concentrations. Overall, the model predicted the concentrations and temporal dynamics of the total sward N concentration satisfactorily for the experimental period. For this component of the system there were few data points over the experimental period to derive any firm conclusions. Nonetheless, the key high and low points in the observed data (particularly in 1998 and 1999) were sufficiently well represented by the model.

The changes in the flux of leaf material from live to standing dead material were also well reproduced by the model (Figure 6c). For the sward, the predicted dynamics of digestibility, metabolisable energy and protein were closely correlated with the observed data. The notable exception was the RH system where the correlation was low (see Table 4) due to a poor (under) prediction of digestibility and metabolisable energy in early summer. This was also the period when there was an under prediction of sheep weight. There were few data points (7–8) of these pasture quality measurements however, which made it difficult to be conclusive about the effects.

3.5. Animals

In the experimental dataset there were a limited number of variables to validate the model output—animal weight and fleece weight. Over the experimental period, the daily variation in live weight of ewes and lambs was moderately well reproduced by the model at Maindample (e.g., Figure 7), but poorly so at Ruffy (see Table 4).

Overall, the weight of the lambs was better predicted than the ewes. For the ewes, the period of poorest agreement with the observed data was during pregnancy, i.e., from mid-January to June. At both locations only mean weights for the mobs were available from the SGS database. Consequently, there was no indication of the variation in the weight of individual animals to assess if the simulations were within reasonable bounds of the data. This mean weight represented, for example at the MM system shown in Figure 6d, the mean weight of all 132 ewes and 180 lambs in the mob.

4. Discussion

In this study, the capacity of a grazing systems model to reproduce the key biophysical changes in a range of complex sheep grazing systems was comprehensively assessed against field data. Although sometimes contentious [34,35], the application of farming systems models to field data is an important process as it points to the strengths of the model and areas that require further work. To the author’s knowledge there are no similar published datasets in Australia, over periods of sufficient length to adequately evaluate grazing systems models. In this light the availability of this dataset it is a rare resource. Especially with highly interconnected grazing models, errors in key components, for example in pasture predictions, will cascade throughout the model. Various authors [36] have cautioned that modification of individual parameter values in complex models is potentially perilous and is most prudently carried against multifaceted datasets. This is what was carried out here. With the detailed SGS experiment dataset, multiple independent checks on the suitability of different model parameters were possible. With minimal modification of parameter values, the model was sufficiently robust, flexible and complex overall to adequately replicate the experimental dataset in terms of the direction of key variables: in response to treatments and extreme climatic events such as, for example, intense rainfall, and heat and cold extremes. Where deviations occurred, reasons could easily be attributed to inconsistencies in the data, assumptions in the modelling or necessary simplifications of complex and heterogeneous systems. Although many of the model parameters are generic in nature, with some minor calibration to local conditions that model represented the complex systems very well overall.

To the author’s knowledge, this is the only such comparison published (in Australia or abroad), between grazing system model predictions and a dataset as such (i.e., soils nutrients, plants, animal, soil water and the water balance). Albeit comprehensive, it was not possible to directly evaluate the functioning of all parts of the model. For example, it was difficult to make conclusive statements about the N cycling aspects of the model as there were no experimental measurements of major system processes such as symbiotic fixation, denitrification, volatilisation, mineralisation and immobilisation, and the size of N pools to assess system N fluxes. However, the functioning of a number of surrogates provides confidence that the model is conceptually sound overall. For example, the soil inorganic N throughout the profile observed in April each year was within the range of the observed data at each site, depth and sampling occasion. Such comparisons (at the beginning of autumn/end of summer) are arguably some of the most rigorous assessments of the system as a whole; soil inorganic N concentrations are potentially influenced by all components of the system. The under-prediction of N in the leachate and soil profile is consistent with the slight over-prediction of pasture dry matter and animal intake at the same times of the year. The closeness between the predicted and measured inorganic soil N in autumn each year is an indication that the conceptual model is sound. This also indicates that key processes are adequately represented in the model, and other variables can be taken with confidence. Assessing the performance of key variables can help to diagnose the effects of problems with data and errors associated with parameter uncertainty. In this regard, the model predictions and the data agreed sufficiently to inform the user that these surrogates are indicators that the overall model works sufficiently well and is conceptually sound. In highly interconnected models, such as the SGS model, errors in plant growth predictions, for example, due to unsatisfactory definition of soil properties, records of daily rainfall or inadequate representation of management, can cascade throughout and compound errors in many other components of the model. For example, the occasional over-prediction of plant growth resulted in an over-prediction of animal weight gain, potentially with subsequent effects on soil N cycling and then on N leaching. Therefore, the key mechanisms involved in N cycling have been captured in the model such as the N dynamics of the legume, sward N dynamics, soil inorganic N dynamics and N in leachate.

For a full discussion of the implications of the treatments on productivity and sustainability the reader is referred to [4,24,25,33]. Reasons for the experimental differences of specific factors, between treatments and due to location, have already been thoroughly investigated and discussed elsewhere and will not be addressed here. Rather, for the remainder of the Discussion, some principles, pitfalls and insights gained when trying to test a complex soil–plant–animal systems model will be addressed. It is intended that this will provide valuable experience and insights for both experimentalists and modellers.

4.1. Using Trends to Ensure Key System Processes Are Well-Represented

The SGS field experiment was conducted on large areas, with the intention that the experiment would provide farming systems information at a scale relevant to farmers [24]. Therefore, the field experiment was not specifically designed to calibrate nor validate the model. Although the large scale did ensure some valuable insights for farmers, the scale of data collection presented some challenges for evaluating the model performance such as some incompatibility in the scales and timing of data collection, issues regarding data quality and the ecological stability (e.g., changing botanical composition due to pasture persistence) of the systems. However, even considering some deficiencies in the quality of the dataset, these can be accepted in the aim for a global validation of the model.

Using the same model across a range of environments and farming systems, the authors of [18] could only explain less than half of the variation in the data during periods of peak pasture growth, such as spring and autumn. The authors referred to the rest as “random variation”. From the analysis in the current study, it has been shown that with better knowledge of critical components such as the temperature responses of the main pasture species groups, legume content and symbiotic fixation and well-defined soil properties, the model’s error can be greatly reduced. However, these are locally important features and may be of less importance in other agroclimatic zones. Although available in the SGS dataset, many of these data are often unavailable and difficult to acquire and therefore variability is to be expected. Even with discrepancy in individual data points, capturing the directions and movement of the key variables is particularly important.

The overall representation of animal growth, metabolism, feed intake, gestation and suckling, as well as fibre production, although simple, was effective in the model overall. Deviations could be explained, for example, in relation to sheep weight prediction in the RH system, where the deviations in early summer could be explained by the associated variations in digestibility and metabolisable energy of the pasture material. Although more parameter-intensive, and more accurate models exist for each of the components, the SGS Model captures the essence of the interacting processes in a systems context [4]. In terms of the greasy fleece weight, weaning weights of the lambs or ewe weight, there were few data points for rigorous validation of the model. Nevertheless, over the experimental period there were no clear systematic biases.

At Ruffy in particular the daily weight by the ewes was sometimes weak but with a much better prediction of weight gain by the lambs. To a small extent the over-prediction of pasture growth at various times had subsequent effects on the predictions of weight gain by the ewes. However, the variable stocking rates and lambing rates were obvious reasons for the differences between observed and predicted data. Additionally, the lambing rate of less than 1.0 lamb per ewe which was recorded at Ruffy was problematic for the model. The lowest lambing rate that can be represented in the model is 1.0 lamb per ewe which is biologically sensible. However, this meant that the simulated weight of all the animals was incorrect at Ruffy.

Overall, the model predictions of pasture biomass production were acceptable but as found by previous authors, the greatest divergence was during peak growth periods when biomass production was at its highest [18,21,22]. This difference at the times of peak biomass production such as during spring may be explained by the lack of representation of reproductive plant growth in the model. Physiological changes that occur during different plant phenological stages are not represented. The trends in pasture growth with maximum production in spring–early summer and minima in winter were adequately reflected by the model, except for the last year of the experiment when there was a shortage of measurements. This is critical to a grazing-focused model as satisfactorily representing pasture dynamics builds the foundation for the functioning of the other component models. The sometimes-poor agreement in the later part of the study was uncertain and may also have been due to inadequacies of the model or due to actual limitations on plant growth due to soil limitations (e.g., aluminium) [24].

4.2. Examples of Challenges Using Field Data Together with Models

Working with variable field data together with biophysical models can be challenging. For example, based on the experimental results, it is difficult to be conclusive about the representation of the N leaching in the model. With closer analysis of the suction cup data, on only 32% of the twenty-two sampling occasions did 50% or more of the number of cups record a sample. There was no occasion when all the cups yielded a water sample. Consequently, it is questionable if the data are suitable for comparisons with the model outputs as it is unclear for example if the differences were due to natural spatial variation in local conditions (e.g., soil properties) or differences due to experimental variation in the effectiveness of the installation of the sampling apparatus or the apparatus itself.

Often the temporal or spatial scale of data collection can often be at odds with those scales modelled. In the SGS study, paddock sizes ranged up to 17.3 hectares. The scale of evaluation of point scale models (i.e., areas larger than experimental plots) as such is rare in the literature due to the issues of spatial heterogeneity in soil properties, for example. Variability is an inherent aspect of all field experimental sites. This can occur with scale due to spatial heterogeneity in soil properties, trees, slope and micro topography, variable areas of accumulation of water, which are also well recognised as influencing N dynamics, pasture responses and soil processes. Averaging data inevitably leads to deviations from model responses in variable environments. This variation could be captured by simulating multiple points within a paddock. SGS has some capacity in this area through the multi-paddock options—knowledge of soil properties, grazing regimes, and hydrological flows with run-on in interacting cells, etc., such as in a fully distributed model.

There were some shortcomings in terms of predicting small surface runoff events, as also found by [37]. Runoff of less than 1 mm/day from the experimental catchment areas was poorly predicted; however, the timing of the large runoff events—when the greatest volumes of runoff occurred—was adequately represented. However, as shown with the data from Ruffy, for the better part the water balance of the entire system was adequately represented and the essence of runoff generating processes were adequately captured by the model for sites where the movement of water via this hydrological pathway was minor.

From the period June 2000 onwards, the botanical composition of the sward changed due to the poor persistence of Cocksfoot pasture as a result of over-grazing and possibly the negative effects of soil acidity [24]. These processes can have a significant impact on the grazing system. As such, changes in botanical composition due to changing plant presence are difficult to replicate in a model. Therefore, models such as the SGS Pasture Model will be limited to some extent in situations where swards are simulated over long time periods where changes can occur in persistence or vigour. This mismatch then between the model and reality confounded the accurate representation of soil moisture in spring and the water balance. The only soil fertility limitation the model deals with is N and representations of the effects of soil P and soil acidity, for example, may potentially improve simulations further. However, increasing the complexity but may also increase further variation thereby compromising an underlying objective of model parsimony.

5. Conclusions

There are an ever-growing group of scientists who use models to understand actual problems/systems and thereby complement or challenge their understanding of complex systems gained from field experimentation. Comparisons with detailed data gathered from a range of sheep farming systems showed that following some minor calibrations to local conditions, the SGS Pasture Model could adequately simulate key water balance, pasture, animal and N cycling processes. Although deviations from the observed data occurred, when interpreted in a systems context they showed that the model realistically represented the systems studied. The major reasons for the divergence between the observed and predicted values were due to the spatial variability of the experimental sites, errors in measurement, challenges associated in representing large heterogeneous sites and the simplified mathematical representation of complex biological systems. When combined with an in-depth understanding of the system, the nature of the pasture systems and performance of key model components, models can reliably provide insights into the long-term biophysical behaviour of a range of agricultural systems. The study demonstrated that such models can be confidently used for investigations of a range of complex grazing situations