Adapting the grassland model BASGRA to simulate yield and nutritive value of whole-crop barley

Ensiling of whole-crop biomass of barley before full maturity is common practice in regions with a short growing season. The developmental stage of barley at harvest can have a large impact on yield and nutritive composition. The relationships between crop growth, environmental conditions and crop management can be described in process-based simulation models. Some models, including the Basic Grassland (BASGRA) model


Introduction
Ensiling the whole biomass fraction, including grain and vegetative fractions, of small-grain cereals before full maturity is a common forage conservation method in regions with a short growing season, such as northern Europe (Wallsten et al., 2010;Rustas et al., 2011;Huuskonen et al., 2017;Randby et al., 2019), or in regions where such crops can be grown during the winter season between two summer crops (Park et al., 2008;Xu et al., 2021).Small-grain cereals are also conserved as whole-crop silage in subtropical and tropical regions with rainy and dry seasons, such as southern Brazil (Leão et al., 2017;Bueno et al., 2018).The developmental stage of the whole crop at harvest can have a large impact on its dry matter (DM) yield and nutritive value (Crovetto et al., 1998;Rustas et al., 2010;Nadeau et al., 2019).The nutritive value, including energy content (Crovetto et al., 1998), sugar, fibre and protein content and fibre digestibility, generally decreases from the heading stage to the dough stage, i.e. the period in which whole crops for ensiling are normally harvested, while the starch content increases (Nadeau, 2007;Wallsten and Hatfield, 2016;Randby et al., 2019).These changes coincide with changes in the ratio of grain to vegetative tissue biomass (grain/straw ratio) (AHDB, 2018) and physiological changes within vegetative and grain tissues (Seog et al., 1993;Khorasani et al., 1997;Gubatz and Shewry, 2010).These changes in nutrient composition in whole-crop barley result in different animal production responses.They also affect how the feed can be used in diets for ruminants differing in production intensity and physiological state (Nadeau, 2007), such as lactating and non-lactating cows.
Growth, development, biomass yield and nutritive composition of barley and other cereal crops are highly dependent on the prevailing weather and soil conditions, and on management practices such as sowing date and seed and fertiliser rate (Slafer et al., 2002).The relationships between these factors are described in different process-based models (Brisson et al., 2003;Jones et al., 2003;Holzworth et al., 2014;Stöckle et al., 2014), which can be used as decision support for farmers and farm advisors.Several crop simulation models have been applied to simulate phenological development, biomass and grain yield of spring and winter barley under different climate conditions (Rötter et al., 2012;Tao et al., 2020).To date, the main priority in barley and other cereal crop modelling studies has been to predict grain yield at full maturity.Less effort has been devoted to predicting aboveground biomass and its underlying components at earlier developmental stages.Moreover, cereal crop models generally do not simulate whole-crop nutritive attributes such as fibre content and digestibility, which are important characteristics.In contrast, there are forage grass models such as CATIMO (Bonesmo and Belanger, 2002b), STICS (Jégo et al., 2013) and the Basic Grassland model (BASGRA) (Höglind et al., 2020) that contain equations for fibre content and digestibility, as the nutritive value of forage grasses is often simulated due to their importance in ruminant feeding.The latter model has been developed primarily for northern European conditions.Its ability to simulate forage grass, including both biomass (Korhonen et al., 2018) and nutritive value attributes such as fibre and protein content and digestibility (Persson et al., 2019), has also been extensively evaluated for this region.Moreover, the BASGRA model includes modules for winter stresses and their effect on winter survival (Höglind et al., 2016), which are processes that are also relevant for winter cereals.In total, this makes BASGRA a suitable candidate to adapt for use in whole-crop cereal simulations.However, the prediction accuracy of both the LINGRA model (Persson et al., 2014), a predecessor to BASGRA, and cereal crop models (Hao et al., 2021) has varied between locations or depending on the geographic distribution of the data that were used for model calibration.Moreover, Angulo et al. (2013), in a study of different calibration approaches of process-based models for five crops in Europe, including winter wheat and winter barley, concluded that regional projections with crop models can be improved if they are calibrated with regional specific data.Therefore, a comparison of different geographical distributions of calibration data could possibly provide information that is useful for further model applications and development of BASGRA for whole-crop barley in northern Europe.
The objectives of this study were to: i) modify the BASGRA model to simulate whole-crop spring barley; ii) evaluate the performance of the modified model against empirical data on DM yield and nutritive value attributes obtained under controlled field experimental conditions; iii) evaluate the importance of the geographic distribution of the calibration data for the model performance; and iv) evaluate DM yield, nutritive value and cutting date under current and projected future climate conditions for three locations in Sweden and four cutting regimes.

Model adaption
The BASGRA model, which was originally developed to simulate DM yield (Höglind et al., 2001;Höglind et al., 2016), crude protein, fibre content and digestibility (Höglind et al., 2020) of the forage grass timothy, was adapted to simulate these four variables for whole-crop barley as a function of weather, soil and crop management.In the forage grass version of BASGRA, plant biomass is divided into four state variables (leaves, stems, roots and reserves).The latter is a pool to which biomass is allocated after allocations to leaves and stems that can later be remobilised to other plant parts (Höglind et al., 2001).Photosynthesis rate is the product of intercepted radiation and photosynthetic light-use efficiency, which takes into account carbon (C) losses through maintenance respiration.The latter depends on CO 2 concentration, through its effect on the photosynthesis rate at light saturation and CO 2 yield per photosynthetically active radiation, as well as air temperature, light intensity and the concentration of Rubisco (Rodriguez et al., 1999;Höglind et al., 2001).The model is based on the source-sink concept, where the source consists of photosynthesising tissue and C reserves, and the sink consists of newly developed above-ground plant parts and roots.Growth respiration is proportional to growth.The soil water is represented by one single layer and is affected by precipitation, runoff, evaporation and plant transpiration.In addition, root growth expands the volume of plant available water.See Höglind et al. (2001) and Höglind et al. (2016) for more details of the above mentioned processes.Simulations can be initialised in a growing sward either with small above ground plant parts early in the spring or after a cut, or with larger above ground plant parts later on in a growth cycle.The soil mineral nitrogen (N) pool that is avalible for plant uptake is, besides plant uptake, affected mineralisation of orgaincally bound N, N fertilisation and deposition, as well as by N losses through leaching and volatilisation processes.The part of the total soil mineral N pool that is avaliable is given by dividing it by a time constant.Given sufficient soil or shoot N available for remobilisation, the N remobilisation is determined by the difference between the actual plant N and the N that would be present if the shoot N followed the light extinction in the sward.Protein concentration is set to N concentration times 6.25.Cell wall content (i.e.fibre content) increases linearly with the plant development and can differ between leaves and stems, while the digestibility of cell wall decreases with the developmental stage.See Höglind et al. (2020) for more details about soil and plant N balance and processes.To simulate C and N dynamics in whole-crop barley in the present study, a spike state variable, equation for allocation of newly photosynthesised C to spikes were added together with cell wall content and digestibility of the spikes.Equations for translocation of C and N from stems and leaves to spikes and associated changes in the sward N canopy were also introduced.The model code can be downloaded from Zenodo (doi: 10.5281/zenodo.8275835).Details of these changes are provided in Sections 2.1.1 to 2.1.5.

Phenology
The parameter PHENGRAINFILL (unitless), which is the developmental stage (on a developmental stage scale from 0 to 1) that sets the start of the grain-filling period during which biomass is allocated to the spike pool, was introduced.PHENGRAINFILL also sets the start of transition of generative tillers, from which no allocation of C to spikes is possible (TILG2; m -2 ), to tillers in which part of the C is allocated to the spikes (TILG3; m -2 ).
Number of tillers, which transitions from the TILG2 to the TILG3 cohort (TILG2G3; m -2 d -1 ) after the start of the grain-filling phase, is driven by a tiller transition rate parameter (RGRTG2G3; d -1 ): Relative death (senescence) rate of leaves due to phenological development (RDRPHEN; d -1 ) is also initiated at the start of the grainfilling phase.

Carbon allocation and translocation to the spike pool
A state variable for spike C (CSPIKE; g C m -2 ) was introduced to divide C between vegetative plant part (leaves and stems) and generative plant tissue (spikes).The increase in biomass in TILG3 tillers is driven by its sink strength (SINK2T; g C tiller -1 d -1 ), which in turn is driven by the ratio between average spike C weight (CSPIKET; g C tiller - 1 ) and maximum possible spike C weight (CSPIKETM; g C tiller -1 ): where SIMAX2T (g C tiller -1 d -1 ) is the sink strength of tillers in the TILG3 cohort at the start of grain filling.The change in sink strength during the grain-filling period is similar to the change in sink strength of tillers in the TILG2 cohort during the elongation phase, which is kept from the original BASGRA model (Höglind et al., 2016).
The direct allocation of C to spikes (ALLOSP; g C m -2 d -1 ) is equal to the smallest of the total allocation of C to sinks (ALLOTOT; g C m -2 d -1 ) and the potential direct allocation rate of C to spike mass (GSPSI; g C m -2 d -1 ).
GSPSI is in turn given by: where TRANRF (unitless) is a transpiration realisation factor (unchanged from the original version of BASGRA), fSPIKE (unitless) is the maximum fraction of newly photosynthesised biomass or biomass mobilised from reserves that can go to the spike pool, NOHARV is an indicator of harvest given a value of 1 when there is no harvest and a value of 0 when the crop is harvested, and YG (g C g C -1 ) is growth yield per unit of C expended.The allocation of C to spikes (ALLOSP) is taken into account in the respiration of shoot growth (RESPGSH; g C m -2 d -1 ) together with the allocation of C to leaves (ALLOLV) and stems (ALLOST): The translocation of C from leaves to spikes (TRANSLLVSP) during the grain-filling period is set as a fraction of the death rate of the leaf mass up to a maximum translocation rate of C from leaves to spikes (TRANSLLVSPMAX; g C m -2 d -1 ).However, the fraction of C in dying leaves that is translocated to spikes is reduced when the daily mean air temperature is above 23 • C or below 10 • C, based on Ritchie et al. (1985) and Pan et al. (2007).TRANSLLVSP is also subtracted from the translocation of C in dying leaves to the soil litter pool.
The translocation of C from stems to spikes is set as a fraction of C in stems per day determined by the relative translocation rate (RSTSPTRANSL; d -1 ), as long as it is below the maximum stem to spike translocation (TRANSLSTSPMAX; g C m -2 d -1 ).As done for leaf to spike C translocation, stem to spike translocation is reduced when the temperature is either above 23 • C or below 10 • C.

Harvest
The harvest equations in the original version of BASGRA were altered to include the tiller cohort with allocation of C and N to spikes (TILG3).

Nitrogen canopy dynamics, allocation and translocation
Plant N dynamics are included in the BASGRA_N version of the model (Höglind et al., 2020), where N content in the plant as a function of light extinction in crop canopy (which regulates plant growth) and N uptake from soil are set to stop at the start of grain filling.Three N state variables (N in stems (NSTEM; g N m -2 ), N in leaves (NLEAF; g N m -2 ) and N in spikes (NSPIKE; g N m -2 )) were added in this study to allow for N translocation from leaves and stems to spikes from the start of the grain-filling period.Protein content is calculated by multiplying the N content in each pool by 6.25.The translocation of N follows the same rules as the C translocation described above, but with separate parameters (Table 1) regulating the translocation rates (Table 2).The translocation of N from leaves to the soil N litter pool is set to stop at the start of grain filling.

Cell wall content and digestibility
Equations were added to describe the digestibility of the spike pool (F_DIGEST_SPIKE; unitless) and the fraction of spike DM that consists of cell wall (F_WALL_SPIKE; g wall g -1 DM).Similarly to the digestibility and cell wall content of other aboveground plant tissue, values of these variables change with phenological development stage.The digestibility of the spike pool increases as the fraction of chaff in the spike pool decreases (Slafer et al., 2002) and the fibre fraction in the grain decreases during the grain-filling period (Seog et al., 1993).

Model evaluation 2.2.1. Crop data
The BASGRA whole-crop model developed was calibrated and validated using data from field trials performed as part of the variety testing programme for cereal crops in Sweden (www.sortval.se).Validation data were excluded from the data that were used for the calibration of the model parameters.Data on DM weight, straw-leaf/spike ratio and nutritive value, including ash, crude protein, neutral detergent fibre (NDF) content and in vitro organic matter digestibility, were obtained for samples of two spring barley (Hordeum vulgare L.) cultivars (cv.Anneli and cv.Judit) taken at Zadoks growth stages 59, 75-77, 83-85 and 89 at Ås (63 • 15'N; 14 • 34'E), Lännäs (63 • 10'N; 17 • 38'E) and   Table 1 Parameters introduced in the BASGRA whole-crop model.Röbäcksdalen (63 • 48'N; 20 • 12'E), Sweden (Fig. 1), in 2019 and 2020.
In vitro organic matter digestibility was transformed to digestible organic matter in the DM (DOMD), according to Givens et al. (2009), by adding the ash content.The variety trial fields were sown in late May or early June, fertilized with between 80 kg N and 100 kg N around sowing, and when needed treated with a chemical herbicide a few weeks after emergence, following normal practices for northern Sweden and the weather conditions in 2019 and 2020.In addition, data on crop DM and nutritive value were taken from previous publications describing specifically designed whole-crop experiments on spring barley, as summarised in Table 3.For the location of the experimental sites see also Fig. 1.In total, those experiments included a wide range of soil and climate conditions, management practices and cultivars and encompassed a large part of the variation under which whole-crop barley is grown in northern Europe.

Weather and soil data
Daily weather data, including minimum and maximum temperatures, accumulated precipitation, global solar radiation, relative air humidity and average wind speed, measured at weather stations near the field experiments were downloaded from the open database of the Swedish Meteorological and Hydrological Institute (SMHI) (https ://www.smhi.se/data/sa-anvander-du-tjansterna-for-oppna-data-1.145103) or from the LantMet network of weather stations administered by the Swedish University of Agricultural Sciences (SLU) (https ://www.slu.se/fakulteter/nj/om-fakulteten/centrumbildningar-och-storre-forskningsplattformar/faltforsk/vader/lantmet/).Data on soil texture fractions were obtained from soil sampling at the trial sites and were used to calculate water-holding capacity characteristics (water content at wilting point, field capacity and saturation) used as input to the BASGRA model, applying methods in the SoilBuild tool (Saxton and Rawls, 2006) in DSSAT software v.4.7 (Hoogenboom et al., 2017).An overview of the climate and soil characteristics and management practices at the field trial locations is provided in Table 3.

Calibration and validation settings
Three model calibrations were carried out to determine model parameter values.Data from all locations were used in the first calibration, data from field trials in northern Sweden sampled in 2003, and 2020 were used in the second calibration, and data from the field trials at Alnarp, Götala, Kungsängen, Lanna and Rådde in central and southern Sweden were used in the third calibration (Table 4).The reason for this division of data was that there are physiological differences between varieties usually grown in northern and southern Sweden, notably differences in temperature requirement for phenological development (www.sortval.se)but also in traits that are related to other parameters whose effects on the evaluated outputs might be better represented if the calibration data is divided according to these geographic regions.In each calibration, the dataset was randomly divided into one subset for model calibration and one subset for model validation (Table 4).A proportion of the total dataset available for each calibration was not used for calibration, but instead used for model validation.These validation subsets consisted of 20%, 19% and 19% of the total data in the whole Sweden, northern Sweden and southern Sweden dataset respectively.
Bayesian calibration techniques as described by Van Oijen et al. (2005) and as previously applied on BASGRA (Höglind et al., 2020) were used.In Bayesian calibration, a prior distribution is updated based on observed data.In the present study, the prior distributions for parameters that were retained from the forage grass version of the BASGRA model were set as equal to the distributions used in a previous calibration against field trial data from northern Europe and Canada (Persson et al., 2019).The prior distributions (Appendix Table 1) of the plant parameters that were added in the BASGRA model version for whole-crop barley were set based on a combination of literature information (Slafer et al., 2002;Seog et al., 1993) (see also Section 2.1 for more information) and results from preliminary calibrations.Initial soil water characteristics were also treated as parameters and their prior distribution was set site-specifically according to soil texture characteristics.Sampling from the posterior distribution was made using the Metropolis algorithm and a chain length of 350000 simulation iterations.The likelihood function (Sivia, 2006) was applied.The model performance was validated using the maximum a posteriori (MAP) parameter values (Appendix Table 2).The simulations were initiated days after sowing using very small initial values of initial leaf area (approximately 0.1 m 2 m -2 ) to represent a newly emerged crop.

Yield and nutritive value under current and future climate conditions
The variation in DM yield and nutritive value of whole-crop barley across climate zones and weather regions within Sweden was evaluated.For this evaluation, we applied the LARS WG tool (Semenov, 2008) v. 6.   et al., 2015).The sets of generated weather data can be considered to represent the variation in the weather under conditions ranging from recent historical to near-future with no stringent mitigation of greenhouse gas emissions.Using 100 years of synthetic weather data for both time periods also meant that we compared the same number of repetitions within each combination of time and location.Dry matter yield and nutritive value attributes were then simulated using the MAP parameter values from the whole Sweden calibration and the generated weather data as input.For each of the three locations (Skara, Umeå and Uppsala), 100 years of simulations were performed using weather data representing 1990-2020 climate conditions and the 2021-2040 HadGEM2-es and RCP4.5 climate.For the latter conditions, the CO 2 concentration in BASGRA was changed from 350 to 435 ppm in accordance with RCP 4.5 (IPCC, 2013).Initialisation date was set to the first day after which mean daily temperature exceeded 5.0 • C for five consecutive days using the same initial leaf area as in the calibrations.Four cutting times, which represented the crop developmental stages at cutting in the experiments from 2019 and 2020 (see Section 2.2.1), were simulated.A fertiliser dose of 100 kg N ha -1 was applied at the start of the simulations in spring in both climate scenarios, at all locations and for all four cutting regimes.

Statistical analysis
The ability of the model to predict nutritive components (crude protein, NDF, DM digestibility) was evaluated using root mean squared error (RMSE): where n is the number of observations and P i and O i are the predicted and observed values for each data pair.The closer the RMSE is to 0, the better the prediction.RMSE was normalised by dividing it by the mean value of the observations.
Relative mean bias error (rMBE) was calculated to provide a measure of the relative magnitude of misprediction (over-or under-prediction): Willmott's index of agreement as a complementary statistical metric was also calculated: where P i ' is P i 'M and O i ' is O i -M, M is the mean of the observed values and d can take a value from 0, which indicates no agreement at all between observed and simulated values, and 1, which indicates complete agreement (Willmott, 1982).
For the 100-year simulations, the difference in mean DM yield, crude protein and NDF content and DOMD between locations, climate projections and cutting times was assessed by one-way analysis of variance (ANOVA) and Fischer's least significant difference (LSD), using PROC GLM in SAS v 9.4 (SAS Institute Inc, 2020).Differences between locations were analysed within climate projections and cutting times, differences between climate projections within locations and cutting times, and differences between cutting times within locations and climate projections.

Model calibration and validation
The prediction accuracy for NDF concentration in whole-crop barley (normalised RMSE 7.58-13.8%)and DOMD (12.3-21.4%)was generally higher than the prediction accuracy for crude protein concentration (17.6-23.2%)and biomass (17.0-27.8%)for the parameter sets based on the three calibrations.Willmott's index of agreement indicated more fluctuating prediction accuracy between crop characteristics and calibrations.The trends in prediction accuracy found in validation of these calibrations, as evaluated by the same metrics, were similar (Tables 5-7).Biomass and crude protein tended to be underpredicted in both calibrations and validations, whereas NDF and DOMD were overpredicted in both calibrations and validations, as evaluated by rMBE.Apart from these differences, it was not possible to discern any trends in prediction accuracy between the calibrations and validations for any of the crop characteristics evaluated (see Tables 5-7 and Figs.2-5 for more details).The prediction accuracy of DOMD was considerably lower for the southern Sweden parameter set than for the other parameters sets, as evaluated by normalised RMSE.For the other variables, there were no clear trends in prediction accuracy between the parameter sets.

Yield and nutritive value under current and future climate conditions
Dry matter yield was significantly lower under the 2021-2040 climate (mean 658.8 g DM m -2 across cutting times and locations) than under the 1990-2020 climate (mean 678.9 g DM m -2 across cutting times and locations) at all four cutting stages and at all three locations (Table 8).The NDF concentration tended to be slightly higher and DOMD slightly lower under the 2021-2040 climate than under the 1990-2020 climate while the effect of climate on crude protein concentration was more variable across locations .Dry matter yield increased, while crude protein concentration, NDF concentration and DOMD decreased with increasing development stage under both climate scenarios and at all locations, i. e. there was a negative relation between dry matter and the three nutritive value components.For both climate scenarios and at all development stages at cutting, dry matter yield was significantly higher at Skara than at the other two locations and mostly significantly higher at Uppsala than at Umeå, across climate scenarios and cutting stages.Crude protein concentration was significantly higher at Uppsala than at Skara and Umeå but mostly not significantly different between the other two locations, whereas NDF   5-7.
T. Persson et al. concentration was significantly higher at Skara and Uppsala than at Umeå at most development stages under both climate scenarios.Digestible organic matter in the DM tended to be highest at Umeå and second highest at Skara for all cutting stages and both climate scenarios.Fig. 9.
All cuttings occurred at earlier dates under the 2021-2040 climate scenario than under the 1990-2020 scenario.The differences in cutting dates between climate scenarios were generally higher at Umeå (on average between 9 and 16 days) than at Skara (on average between 7 and 9 days) and Uppsala (on average between 8 and 9 days) and also tended to be higher at later developmental stages than at earlier stages.Moreover, the cutting date varied less under the 2021-2040 climate scenario than under the 1990-2020 climate scenarios, especially at Umeå (Fig. 5).

Model performance
In this study the BASGRA model, which was originally developed for forage grass, was adjusted to simulate yield and nutritive value attributes of whole-crop barley under a range of environmental conditions and genetic diversity.There are several possible reasons for the fact that the prediction accuracy was generally higher for NDF, and partly also for crude protein and DM yield, than for DOMD.First, the high prediction accuracy of NDF could be related to lower genetic variability in the plant   5-7.
T. Persson et al. functions that regulate this nutritive value attribute across cultivars than in those regulating biomass, crude protein content and digestibility.The overall higher prediction accuracy of NDF than of crude protein is in line with findings in a previous evaluation of the BASGRA CATIMO and STICS forage grass models for timothy grass (Persson et al., 2019;Bonesmo and Belanger, 2002a;Bonesmo and Belanger, 2002b;Jégo et al., 2013), indicating that prediction accuracy for NDF is rather stable across grass ley and cereal crops, while the low prediction accuracy of digestibility differs from findings for timothy grass in the evaluation of the same forage grass models (Persson et al., 2019).There was greater variation of biomass within many of the datasets used than within the measured nutritive value attributes (data not shown), suggesting that the former observations were associated with greater uncertainty, which may explain some of the biomass prediction error.Possible measurement errors could also have contributed to the high variability in observed biomass since, especially in uneven stands, there is arguably a higher risk of obtaining non-representative biomass weight samples than nutritive value attribute samples.A greater variation in observed biomass than in observed nutritive value has also been found in previous studies of forage grass (Bélanger et al., 2008;Nissinen et al., 2010).

Regional differences
That the prediction accuracy in the northern Sweden calibration was not better than that in the calibration for the whole country could be  because to physiological differences within the former dataset have a larger impact on the model performance than physiological differences between cultivars at the northern and southern locations or environmental differences between geographic regions that are not fully mechanistically accounted for in the BASGRA model.The fact that there was one two-row barley cultivar (Anneli) and two six-row barley cultivars (Judit and Olsok) in the northern Sweden dataset with different physiological traits supports this argumentation.More data than what was available for this study would be needed to test if the effect of cultivar-specific calibration could improve prediction accuracy.The southern Sweden calibration dataset consisted of a wide range of cultivars with different properties, which could likewise explain why this calibration did not improve the prediction accuracy compared with the calibration for the whole country.The lack of effect of geographical region used in calibration on model performance differs from findings in previous model calibrations for wheat (Hao et al., 2021) and forage grass (Persson et al., 2014), in which the yield predictions increased after site-specific factors were taken into account.These differences in regional effect between crops could be due to a higher geographic variability in stressors and other factors affecting winter wheat and perennial forage grasses but not whole crop spring barley, notably winter stress, which is not described with a very high level of details in most cereal crop (Hochman et al., 2013;Bergjord Olsen et al., 2018) and forage grass models (Kipling et al., 2016).Overall, this indicates that regional specific whole crop model calibrations taking into account variations in climate, soil and other environmental factors should be combined with other measures.These include division of datasets into specific cultivars, climate and soil conditions, requiring larger sets of cultivar and site-specific data than were used in this study.Moreover, both protein and fibre concentration, and digestibility of stems and leaves (Bonesmo and Belanger, 2002b;Nissinen et al., 2010) have been shown to vary considerably between environmental conditions and developmental stages in other grass species than barley.Hence, nutritive value specifically for stems and leaves, as well as for spikes could be useful in understanding the effect of C and N allocation and translocation on the change in nutritive value during plant development, especially the reason for overprediction of DOMD and underprediction of crude protein content.

Practical applications
Apart from DOMD, the prediction accuracy was sufficiently high for the modified model to be useful in evaluating the performance and composition of whole-crop barley under different production conditions in Sweden or other countries with similar climate, soil and crop management practices.The higher prediction accuracy for NDF than for crude protein and biomass could be important in applications of the model in e. g. forecasting systems for forage production.In ruminant feeding regimes, the fibre component, which gives structure to the diet (Banakar et al., 2018), is usually mainly supplied by forage, e. g. whole-crop barley.Still, at advanced developmental stages, a reduction of the NDF concentration promotes higher intake, which to some extent can compensate for a relatively low energy content (Rustas et al., 2010).The protein requirement of cattle and other ruminants is largely covered by feed components such as grass-legume silage (Castro-Montoya and Dickhoefer, 2020; Chowdhury et al., 2023) and concentrate (Wilkinson, 2011;Erickson and Kalscheur, 2020), which is also true in feeding regimes that include whole crop components, especially when the whole crop is cut at a late developmental stage (Hetta et al., 2012;Wallsten et al., 2009).

Climate impact on whole crop production
The rather small differences in NDF, DOMD, crude protein and biomass between locations and climate conditions in the 100-year simulations suggest that production of whole-crop barley is rather stable under the conditions studied and under regional climate conditions representing the near future.However, the small negative impact of climate change on dry matter yield differs from simulated positive effects of climate change on forage grass yield, largely as an effect of longer growing seasons in the same geographic region (Höglind et al., 2013).The simulated earlier cutting dates and lower cutting date variability under the 2021-2040 climate than under the 1990-2020 also indicate that the changing climate could allow for more cropping flexibility.This flexibility gives a longer time for field preparation and sowing in the spring, which can be a particular bottleneck in high-latitude regions with relatively high precipitation (Kolberg et al., 2019).It also gives a longer time for winter crop sowing in autumn, provided that projected future increases in precipitation and higher frequency of extreme precipitation events (Lehtonen et al., 2014) do not impair the conditions for seedbed preparation and sowing.The slightly lower yield, which possibly could be due to a shorter period for carbon assimilation which in turn can be a result of faster phenological development (Tao et al., 2008;Fatima et al., 2020) may also to some extent limit future whole crop production.

Further research priorities
This study adds to previous findings on the sensitivity of cereals and other agricultural crops to weather and climate variability, which have focused mainly on yield quantity (Tripathi et al., 2016;Wang et al., 2018) and to some extent on N and protein (Asseng et al., 2019), but have largely excluded other quality aspects.However, there are also reasons to treat the results from our study with some caution.Climate projections differing from that considered here might have generated different results given the large impact of climate projections and global climate models found on other crops such as grain cereals (Challinor et al., 2009;Asseng et al., 2013) and forage grass (Graux et al., 2013;Höglind et al., 2013).The deviation between simulated and observed variables, in particular the overprediction of digestibility, suggests a need for further model development before future evaluations of climate impact, e.g.including a wider range of climate change projection, on whole-crop barley production.This could also include sensitivity tests to dry and wet spells, as adjusted by LARS-WG or another weather generator, to further emphasize risks and preparation for worse case scenarios for seed bed preparation, sowing and harvest.Such efforts could also take into account a wider geographic selection of locations, soil properties and management practices, notably N fertilisation, which are representative of Sweden and other regions where whole crop barley is commonly grown or has a potential to become an important animal feed component.Such an approach could improve the prediction accuracy and improve understanding of changes in whole-crop production due to climate change and other environmental variation in an applied perspective.Further research tasks to improve model prediction accuracy could include calibrations based on cultivar differences (e.g. between two-row and six-row barley cultivars) and detailed studies about N and C translocation between vegetative biomass and spikes and its link to changes in fibre concentration and digestibility during crop development.Such studies could be accompanied by future scenario analyses and piecemeal implementation of the model in decision support systems for famers and agricultural advisors.

Conclusions
The BASGRA model for forage grass was adapted to simulate whole-crop barley, notably by adding a spike pool, equations for C and N allocation to the spike pool and equations for C and N translocation from vegetative plant parts to spikes.Model calibrations and validations using field trial data from across Sweden, and across crop developmental stages from late anthesis to hard dough, generally showed better prediction accuracy of nutritive value attributes, especially NDF, than of DM yield.Except in the case of DOMD, prediction accuracy was sufficiently high to allow the model to be used for evaluating the performance of whole-crop barley under different production conditions in Sweden.Region-specific calibrations for southern and northern Sweden did not change the pattern of prediction accuracy.An evaluation of climate change effects showed lower DM yield, earlier harvest and slightly higher NDF concentration on average for 100 years of generated weather data representing near-future climate conditions (2021-40) than data representing historical climate conditions (1990-2020) at three locations in Sweden (Skara, Umeå and Uppsala), using the model version calibrated against the dataset for the whole country.Further model development, using additional observed data, is needed to improve the prediction accuracy for digestibility.

First,
Swedish Meteorological and Hydrological Institute (SMHI) daily data on minimum and maximum air temperature, precipitation, and solar radiation for Skara (58 • 23'N;13 • 27' E; 118 m asl), Umea (63 • 47'N; 20 • 17'E; 7 m asl) and Uppsala (59 • 53'N; 17 • 36'E; 17 m asl) and the period 1991-2022 were downloaded from the website of the Field Research Unit (Fältforsk) at SLU (https://www.slu.se/fakulteter/nj/om-fakulteten/centrumbildningar-och-storre-forskningsplattfor mar/faltforsk/vader/lantmet/).Second, site information and data files were created according to the manual included in LARS-WG v. 6.Third, a site analysis to analyse dry and wet spells, and cold and hot spells for the three sites.Finally, for each of the three locations, 100 years of synthetic daily data on minimum and maximum air temperature, precipitation, and solar radiation were generated for the baseline period, and for the time period 2021-2040 and the Global climate model Hadley Centre Global Environment Model version 2 (HadGEM2-es) and the representative concentration pathway (RCP 4.5).The latter represents a moderate greenhouse gas emission scenario, which includes an average global warming of 1.5 • C (IPCC, 2013) and approximately 2 • C in Northern Europe (Hansen-Bauer et al., 2015) compared to the temperatures during the late 20th century and early 21st century.The climate changes for northern Europe associated with RCP 4.5 also include increases in annual precipitation until 2050 (Hansen-Bauer

Fig. 2 .
Fig. 2. Observed biomass plotted against simulated biomass for the three a) model calibrations and b) model validations.Statistics on calibrations and validations are presented in Tables 5-7.

Fig. 3 .
Fig. 3. Observed crude protein plotted against simulated crude protein for the three a) model calibrations and b) model validations.Statistics on calibrations and validations are presented in Tables 5-7.

Fig. 4 .
Fig. 4. Observed neutral detergent fibre (NDF) plotted against simulated NDF for the three a) model calibrations and b) model validations.Statistics on calibrations and validations are presented in Tables 5-7.

Fig. 5 .
Fig. 5. Observed digestible organic matter in the dry matter (DOMD)plotted against simulated DOMD for the three a) model calibrations and b) model validations.Statistics on calibrations and validations are presented in Tables 5-7.

Table 2
Variables introduced in the BASGRA whole-crop model.

Table 3
Climate, soil characteristics and management practices at the experimental locations.

Table 4
Calibration and validation datasets.

Table 5
Statistics for the whole Sweden calibration.Observed and simulated means, root mean squared error (RMSE), normalised RMSE, relative mean bias error (rMBE) and Willmott's index of agreement (d-index) for biomass crude protein (CP) concentration, neutral detergent fibre (NDF) concentration and digestible organic matter in the dry matter (DOMD).

Table 6
Statistics for the northern Sweden calibration.Observed and simulated means, root mean squared error (RMSE), normalised RMSE, relative mean bias error (rMBE) and Willmott's index of agreement (d-index) for biomass crude protein (CP) concentration, neutral detergent fibre (NDF) concentration and digestible organic matter in the dry matter (DOMD).

Table 7
Statistics for the southern Sweden calibration.Observed and simulated means, root mean squared error (RMSE), normalised RMSE, relative mean bias error (rMBE) and Willmott's index of agreement (d-index) for biomass crude protein (CP) concentration, neutral detergent fibre (NDF) concentration and digestible organic matter in the dry matter (DOMD).

Table 8
Effect of climate projections on dry matter yield (g DM m -2 ) of whole-crop barley cut at different developmental stages.