Assessment of Nitrogen Uptake and Biological Nitrogen Fixation Responses of Soybean to Nitrogen Fertiliser with SPACSYS

Chemical fertiliser nitrogen addition will inhibit biological nitrogen fixation (BNF) for soybean (Glycine max [L.] Merr) growth. The optimal balance of these two nitrogen input sources has been a key issue for sustainable development in Northeast China. We used the data collected from a four-year experiment with varied irrigation and fertiliser treatments from 2007 to 2010 to evaluate the SPACSYS (Soil-Plant-Atmosphere Continuum SYStem) model. The validated model was run to investigate the responses to different management practices in seed yield, BNF, protein yield and soil nitrogen budgets. Scenario testing showed average yield increase of 2.4–5.2% with additional 50–100 kg N/ha application. Irrigation at the reproductive stage improved seed yield in drier years with an increase of 12–33% compared with the rain-fed treatment. BNF was suppressed by fertiliser nitrogen application and drought stress with a decrease of 6–33% and 8–34%, respectively. The average nitrogen budget without fertilization indicated a deficit of 39 kg N/ha. To attain higher seed yield, applying fertiliser at 25–30 and 15–20 kg N/ha before sowing is advised in drier and wetter years, respectively. To achieve a higher seed nitrogen content, an application rate of 55–60 and 45–50 kg N/ha is recommended for drier and wetter years, respectively.


Introduction
The soybean (Glycine max [L.] Merr) is the fourth largest crop in China and one of the main crops in the Northeast China plain (NEP). The local farmers' interest in growing soybean has steadily increased since the 1980s [1,2], and total sown area and production in the NEP accounted for more than 30% of the national total in 2015 [2]. Average yields were similar for both the NEP and the whole country before the 1980s but since then have been 30% higher for the NEP [2,3]. This increase is mainly attributed to cultivar improvement and better management practices, including nitrogen (N) fertilization and tillage [3,4].
To maintain higher soybean yields, N should be supplied at the right time with the right amount. Other than N provided by soil mineralization, biological N 2 fixation (BNF) and N fertilizer (either chemical or manure) are the two important sources to meet soybean N demand. BNF normally starts at the third node (V3, three nodes on the main stem with fully developed leaves beginning with the unifoliolate nodes) stage [5][6][7] with a lower rate until the sixth node (V6) stage [5][6][7]. It reaches a maximum from pod setting to the beginning of the seed filling stage, which coincides with the peak N demand for protein synthesis and then decreases [7][8][9][10]. If N supply by BNF and soil mineralization is insufficient to satisfy crop demand, additional N should be supplied, otherwise photosynthesis could Figure 1. Ten-day average temperature (A), accumulated precipitation (B) and total sunshine hours (C) during the soybean growing season from 2007 to 2010 and long-term average at the Gongzhuling experimental site. 1 represents the first ten days, 2 represents the middle ten days and 3 represents the last ten (eleven) days. The weather conditions across the growing season were quite different in each year. Average temperatures during the growing season for each year were slightly higher than the long-term average, especially for the vegetative phase (  Ten-day average temperature (A), accumulated precipitation (B) and total sunshine hours (C) during the soybean growing season from 2007 to 2010 and long-term average at the Gongzhuling experimental site. 1 represents the first ten days, 2 represents the middle ten days and 3 represents the last ten (eleven) days. The weather conditions across the growing season were quite different in each year. Average temperatures during the growing season for each year were slightly higher than the long-term average, especially for the vegetative phase  the long-term average, while those for the 2008, 2009 and 2010 growing seasons were lower, with  values of 852, 918 and 794 h, respectively. However, sunshine hours for the seed filling period in 2007  (281 h), 2008 (238 h) and 2009 (212 h) were higher than the long-term average (210 h), and lower in 2010 (157 h). In summary, 2007 was drier with the most sunshine hours, 2008 wetter but with a slight drought stress at the reproductive stage and fewer sunshine hours, 2009 drier with fewer sunshine hours and 2010 wetter with the least sunshine hours (Figure 1).

Experimental Design and Data Collection
The experiment included two factors (fertiliser and irrigation) using a split-plot design with a row spacing of 0. The dates for the main phenological stages including emergence, initiation of flowering (R1), full flowering (R2), pod setting (R4), seed filling (R6) and maturity (R8) were recorded following the description by Fehr et al. [23] Three plants were randomly selected in each plot at these stages, and the leaf area index (LAI) and dry biomass of roots, leaves, stems and pods if present was measured in each growing season. Seed yield (dry matter) at harvest was calculated based on a 2 m 2 quadrat sample. LAI was estimated by the disc method [24]. Ten leaves were randomly chosen from the three plants and all of them were perforated once in the central nervure. Root biomass was measured by the root core method to a depth of 100 cm, with a 10 cm core diameter [25]. Roots were picked out and both leaf discs and roots were dried in a forced aeration oven at 70 • C until constant mass was achieved. Soil volumetric moisture content was measured using soil cores to the depth of 60 cm, divided into 4 layers with the top two layers of 10 cm each and the lower two of 20 cm each. Three cores per plot were taken randomly. The measurements were taken seven to eleven times over the growing season. The N content in each organ (roots, leaves, stems and seeds) was measured using the Kjeldahl method at each growth stage in 2008 (only for treatments with irrigation at different fertiliser application rates) and 2009 (for all the treatments).
Daily meteorological data for the site (maximum and minimum temperatures, sunshine hours, precipitation, relative humidity and wind speed) were downloaded from the China Meteorological Administration [26]. Daily solar radiation was estimated from daily sunshine hours [27].

The SPACSYS Model
The SPACSYS (Soil-Plant-Atmosphere Continuum SYStem) model is described fully by Wu et al. [22] The generic crop growth and development component describes phenological development, daily photosynthesis rate, root penetration and growth, N uptake, partitioning and translocation of photosynthate and absorbed N and the senescence of leaves, stems and roots. A BNF rate is calculated according to one of two algorithms, based on either root biomass [28] or above-ground biomass [29]. Nitrogen cycling is coupled with C cycling and comprises external N input (deposition, chemical fertiliser and manure), organic matter decomposition, nitrification, denitrification and N Sustainability 2020, 12, 5921 5 of 17 losses through leaching and surface runoff. Organic matter (OM) decomposition flow includes fresh litter, dissolved OM, humus and microbial OM. The model algorithms are described in detail elsewhere [22,30,31], but we present here the specific calculations for the BNF rate (F fix ) based on root biomass: where W root is root biomass; f nodule is the fraction of nodule in root biomass; N con is soil mineral N content, which may inhibit the process of N 2 fixation at higher values [27,32,33]; and f t and f w are the temperature and water impact factors, respectively: where T fixmin is the minimum temperature below which fixation ceases; T fixoptl and T fixoptu define the optimum temperature range; T fixmax is the maximum temperature above which fixation stops; W max is maximum available water capacity; and T and W a are soil temperature and available water content, respectively.

Model Calibration and Validation
A subset of data for seed yield, dynamics of LAI, dry matter in different plant organs and N content from the F0 × W0 treatment over the period between 2008 and 2010 was used to calibrate the SPACSYS model. The remaining data were used for model validation. The plant genetic parameter values that lead to minimum bias at the calibration stage were used for validation and scenario simulation (see next sub-section). The model was run from 2005 to reduce the side-effects of the configured initial conditions. The calibrated parameters used in this study are shown in Table 2. A combination of 4 statistical indicators was used to evaluate model performance: (1) the correlation for determination (R 2 )-evaluating the percentage of variation in observations explained by the model; (2) the relative root mean square error (nRMSE%)-quantifying the relative magnitude of error; (3) the modelling efficiency (EF)-determining whether the modelling output matches the observed data (EF > 0 is the condition to achieve "good fitting results" between simulated and observed data); and (4) the index of agreement (D)-reflecting the degree to which the observed variate is accurately estimated by the simulated variate [34]. EF and D are calculated as: where n is the total number of samplings; O i and O are the ith observation value at sampling i and the average of the observation values, respectively; S i and S are the ith simulation value and the average of the simulation values, respectively and "||" is an absolute calculation.

Simulation Scenarios Design
In order to search for optimum agronomic management practices for maintaining higher soybean yields while maximizing BNF, we designed various scenarios regarding fertiliser N application and irrigation to simulate their effects on soybean yield, BNF and seed N content. Five fertiliser application strategies (application before sowing, BS; application at initial flowering stage, R1; application at pod initiation stage, R3; application before sowing and N top dressing at initial flowering stage with split amount of 7:3, BR1; application before sowing and N top dressing at pod initiation stage with split amount of 7:3, BR3) with various application rates (0, 50 and 100 kg N/ha) and four irrigation schedules (rain-fed, W0; irrigating before flowering stage, Wr; irrigating at the R3 stage, W1; irrigating at both R3 and R5 stage with a doubled total irrigation amount, W2) were included, giving 44 scenarios in total. For each irrigation event, 40 mm of water was applied [35]. Further simulations were designed for N fertiliser rates between 0 and 100 kg N/ha with an increment of 10 kg N/ha applied before sowing to study the optimum fertiliser N rate for achieving a higher seed yield, and the same rates applied at the reproductive stage to achieve a higher seed N in the NEP.

Nitrogen Budget Calculation
A soil N budget during the growing season was calculated based on the outputs from the simulations for the different scenarios: (6) where N fert is N input from chemical fertiliser and manure; N depo is N input from atmospheric deposition; N fix is N input from BNF; N offtake is N removed through harvested seeds and residue; and N loss and N deni are N losses through leaching (below 1.5 m soil depth) and surface runoff, and denitrification, respectively. When N bud-soil < 0, it can be inferred that the soil is a net source of N, otherwise it is a net sink.

Model Calibration and Validation for Plant Growth, Soil Water Content and Soil Temperature
Validation of the calibrated model output for crop dry matter accumulation, N content in various organs and LAI at different stages, soil temperature and moisture during the growing seasons for the F0 x W0 treatment is shown in Figures 2-4. Overall, the simulated results were at an acceptable level, with simulated trends and patterns the same as the observed ones ( Table 3). The model captured the trend in leaf biomass and LAI well, with R 2 > 0.9 and EF > 0.8 ( Figure 2B-F, Table 3). The simulation bias for stem biomass was larger in the first three stages, with a better fit between the simulated and observed values in later stages ( Figure 2G-I). Root biomass was generally well simulated with a little overestimation in the early stages ( Figure 2J-L). Simulated N content in different organs agreed with measured values in most cases, especially for leaf and seed, with nRMSE% < 20% ( Figure 3, Table 3).     The model captured the peaks and troughs in soil temperatures reasonably well ( Figure 4A,B). Specifically, the model reproduced the fluctuations in the 10 cm soil temperature better than those for 20 cm, for which it slightly under-predicted during the vegetative stage. Dynamics of soil water content in the soil profile were captured by the model with nRMSE% < 15%. The model performed better for deeper soil layers (data not shown) than for the top soil layers. The simulated values for the top 10 cm soil layer matched the observed ones well, but slightly under-predicted against 2008 observations ( Figure 4C). Statistical analysis for the comparison between simulated and measured values for validation is presented in Table 3. A comparison between simulated and measured seed yield for all treatments is shown in Figure 5. Both estimated yield and seed N content matched observations very well, with nRMSE% of 12.1% and 7.7%, respectively. Although leaf dry matter, leaf N content and LAI showed  Table 3. Statistical analysis of simulated soil water content, leaf area index, dry matter and nitrogen (N) content of leaves, stems roots and seeds and seed yield. (R 2 ) the correlation for determination; (nRMSE%) the relative root mean square error; (EF) the modelling efficiency; (D) the index of agreement. The model captured the peaks and troughs in soil temperatures reasonably well ( Figure 4A,B). Specifically, the model reproduced the fluctuations in the 10 cm soil temperature better than those for 20 cm, for which it slightly under-predicted during the vegetative stage. Dynamics of soil water content in the soil profile were captured by the model with nRMSE% < 15%. The model performed better for deeper soil layers (data not shown) than for the top soil layers. The simulated values for the top 10 cm soil layer matched the observed ones well, but slightly under-predicted against 2008 observations ( Figure 4C).
Statistical analysis for the comparison between simulated and measured values for validation is presented in Table 3. A comparison between simulated and measured seed yield for all treatments is shown in Figure 5. Both estimated yield and seed N content matched observations very well, with nRMSE% of 12.1% and 7.7%, respectively. Although leaf dry matter, leaf N content and LAI showed large relative errors, the simulated output matched observed data very well, with EF and R 2 close to 1. Similarly, the general trend and pattern of the simulated root biomass matched the observations. Simulation for stem biomass and stem N content was less good, with nRMSE% > 30%, but R 2 > 0.5 and EF > 0. From these comparisons, we concluded that the simulated outputs were acceptable and that the SPACSYS model could be used for further scenario analysis.

Response of Seed Yield to Simulated Field Management Strategies
Soybean seed yield varied from 1970 to 3400 kg/ha under the different simulated field management strategies (Table 4). Yield increased with N fertiliser application rate and irrigation, but the rate of increase differed under dry and wet growing conditions. The highest irrigation use efficiency and higher yield increase occurred in drier years. For example, yield increased by 3.9-15.2% and 20.9-33.3%, compared with yield from W0 in 2007 and 2009, respectively, but there were no differences between the irrigated and rain-fed treatments for 2008 and 2010. Notably, compared with the rain-fed treatment, yield increased substantially by 12% for treatment W1 but only by 4% for treatment Wr in 2007, with a further 2% increase for treatment W2 compared with that for treatment W1 in the same year. However, yield increased by 26-29% for treatment Wr in 2009, and a further 4-5% improvement for treatment W2. Yield increased by up to 9% for an increase in N application from 50 to 100 kg N/ha, a smaller response to that from increased irrigation. When fertiliser was applied as a single application, seed yield was higher with basal application than for application during the reproductive stages. These results demonstrated that N supply was not the main constraint to yield increase in this region but that climatic conditions were. Irrigation applied during the reproductive stage resulted in higher yield regardless of whether there was drought stress. Sustainability 2020, 12, x FOR PEER REVIEW 10 of 18

Response of Seed Yield to Simulated Field Management Strategies
Soybean seed yield varied from 1970 to 3400 kg/ha under the different simulated field management strategies (Table 4). Yield increased with N fertiliser application rate and irrigation, but the rate of increase differed under dry and wet growing conditions. The highest irrigation use efficiency and higher yield increase occurred in drier years. For example, yield increased by 3.9%-15.2% and 20.9%-33.3%, compared with yield from W0 in 2007 and 2009, respectively, but there were no differences between the irrigated and rain-fed treatments for 2008 and 2010. Notably, compared with the rain-fed treatment, yield increased substantially by 12% for treatment W1 but only by 4% for treatment Wr in 2007, with a further 2% increase for treatment W2 compared with that for treatment W1 in the same year. However, yield increased by 26%-29% for treatment Wr in 2009, and a further 4%-5% improvement for treatment W2. Yield increased by up to 9% for an increase in N application from 50 to 100 kg N/ha, a smaller response to that from increased irrigation. When fertiliser was applied as a single application, seed yield was higher with basal application than for application during the reproductive stages. These results demonstrated that N supply was not the main constraint to yield increase in this region but that climatic conditions were. Irrigation applied during the reproductive stage resulted in higher yield regardless of whether there was drought stress.

Response of BNF to Field Management Strategies
Nitrogen fixation responded positively to the increase in irrigation but negatively to the increase in fertiliser N application and delays in application timing.

Soil Mineral N Budget under Various Field Management Strategies
The soil mineral N budget over the soybean growing season was negative when no fertiliser was applied, with a range between −3.5 and −58.2 kg N/ha and an average value of −39.0 kg N/ha ( Figure 6). It increased with N fertiliser application; a treatment of 100 kg N/ha leads to an average increase of 28. Atmospheric deposition over the growing season ranged from 20.6 to 28.9 kg N/ha, responsible for 9-22% of the total N input. From the simulations, N loss through denitrification ranged between 1.5 and 7.9 kg N/ha. The loss through leaching and surface runoff went from nil to 37 kg N/ha and became larger at the higher application rate with additional irrigation application, especially in 2010, when N losses accounted for 10.9-13.2% of the total input under the W0 treatment and for 15.1-18.3% under the W2 treatment. Meanwhile, it was between 10.9-16.8% with incorporation of basal fertilization, with an additional 1.0-1.5% loss if a single fertiliser application was made at the reproductive stages.
rain-fed treatments. Response of seed N content to fertiliser N application rate might not be economical, with an increase of 7.6-23.7 kg N/ha when doubling the application rate from 50 to 100 kg N/ha. Seed N content was the least when fertiliser was applied as a single dose at the reproductive stage, especially at R3. Basal application only resulted in a higher seed N content compared with that applied at R3 alone, with an average increase of 2.8, 1.

Soil Mineral N Budget under Various Field Management Strategies
The soil mineral N budget over the soybean growing season was negative when no fertiliser was applied, with a range between −3.5 and −58.2 kg N/ha and an average value of −39.0 kg N/ha ( Figure  6). It increased with N fertiliser application; a treatment of 100 kg N/ha leads to an average increase of 28. Atmospheric deposition over the growing season ranged from 20.6 to 28.9 kg N/ha, responsible for 9%-22% of the total N input. From the simulations, N loss through denitrification ranged between 1.5 and 7.9 kg N/ha. The loss through leaching and surface runoff went from nil to 37 kg N/ha and became larger at the higher application rate with additional irrigation application, especially in 2010, when N losses accounted for 10.9%-13.2% of the total input under the W0 treatment and for 15.1%-18.3% under the W2 treatment. Meanwhile, it was between 10.9%-16.8% with incorporation of basal fertilization, with an additional 1.0%-1.5% loss if a single fertiliser application was made at the reproductive stages.

Recommendations for Fertiliser N Application Rate in Northeast China
In general, irrigation at the reproductive stage increases seed yield, BNF and seed N in drier years, while the rain-fed system is better in wetter years. Fertiliser applied before sowing (BS, BR1, BR3) can improve seed yield and N content, but when applied at the reproductive phase resulted in an additional reduction in BNF (Table 4). Biological N fixation is generally regarded as an environmentally friendly source of N. However, soil N budgets were negative if no fertiliser was applied during the growing season ( Figure 6), indicating that the system was unsustainable. A certain amount of fertiliser N as basal application could, therefore, be beneficial for BNF compared with other application schedules. However, higher rates constrained BNF. The simulation results indicated that the appropriate amount (the intersection of the two lines in Figure 7) of N fertiliser for the rain-fed treatment could lead to yield increase without reducing BNF by too much (left panels in Figure 7). A fertiliser application rate of 31.5, 19.4, 30.5 and 16.3  indicated that the appropriate amount (the intersection of the two lines in Figure 7) of N fertiliser for the rain-fed treatment could lead to yield increase without reducing BNF by too much (left panels in Figure 7). A fertiliser application rate of 31.5, 19.4, 30.5 and 16.3 kg N/ha was appropriate for 2007, 2008, 2009 and 2010, respectively. With irrigation at the reproductive stage, the appropriate fertiliser application rate was 26.2, 21, 29.6 and 16.6 kg N/ha for 2007 to 2010, respectively (data not shown). Combining all simulations, it is suggested that 25-30 kg N/ha can be applied before sowing during drier years with/without irrigation, and 15-20 kg N/ha can be applied in wetter years.  Higher seed N content is used as an indicator of seed quality. From the simulations, fertiliser applied at the reproductive stage only resulted in lower seed N content compared to the strategies in which a basal application was included (Table 4). An appropriate fertiliser application rate of 57.5, 47.9, 58.5 and 41.9 kg N/ha before sowing was advised for 2007, 2008, 2009 and 2010, respectively, under rain-fed conditions (right panel in Figure 7) and 55.9, 48.7, 58.0 and 42.1 kg N/ha under irrigation (data not shown). In summary, the simulated results indicated that an application rate of 55-60 kg N/ha was suitable to achieve a higher seed N content in drier years, and 40-50 kg N/ha in wetter years. A higher N application rate is more important to maintain higher seed N content than for yield.

Discussion
The SPACSYS model was validated against the field experiment data from 2007-2010 in Northeast China. Results showed that after calibration it could effectively reproduce the observed dynamics of soybean growth in terms of biomass and N content in leaves, roots and seeds, final yield, soil temperature and soil water content (Table 3). However, larger discrepancies still existed for stem biomass and N content, which might be caused by simplified photosynthate partitioning coefficients to various organs in the model, but these discrepancies were less important in the context of our study. Such discrepancies are not uncommon [36], and in an inter-comparison of five soybean simulation models, Battisti et al. (2017) reported that results for crop growth under different climate conditions were often unsatisfying [37]. As a legume crop, it is particularly important to consider the ability to model BNF in soybean. Quantitative assessment of this was not included in recent published simulation results for North China that mainly focused on final yield, soil water and evapotranspiration [21,38]. The inclusion of the simulation of the BNF rate in SPACSYS, using the root-based algorithm (with the model giving reliable estimation of root biomass, Table 3), therefore, represents an important progression in our ability to model soybean production and, particularly, how it is influenced by different irrigation and N fertiliser application strategies.
From the model simulations, soybean yield significantly increased with irrigation during the reproductive stage under dry conditions, consistent with previous observations that a single irrigation event at the reproductive stage increased yield by 30% [4]. However, the effect of N fertiliser application on yield was not as obvious compared to irrigation for all the experimental years (yield increase < 10% with a fertiliser application rate up to 100 kg N/ha). Previous experimental results have shown a similar yield increase with fertilization [18,33]. Seed N content increased in response to N application, consistent with field observations made using the 15 N natural abundance method [39], and between 179-211 kg N/ha was accumulated in seed in the wetter years, but only 60-70% of this amount was achieved in 2009 without irrigation ( Table 4). The effect of N fertiliser rate on soybean yield may vary with soil fertility, where differing amounts of mineral N can become available through mineralization [15,33]. In our study, we suggest that climate conditions were the major constraint on yield. The soil in the study area was fertile with a high organic matter content, where N supply from mineralization (with a range from 88 to 157 kg N/ha from the simulations) and BNF would be sufficient for soybean growth with low yield potential [10], meaning that weather conditions, especially photosynthetic active radiation (PAR) and rainfall, are the main driving variables for seed yield [3]. Soybean has a high water requirement between growth stages R1 to R5 [35]. By alleviating drought stress during this period, a 20-30% yield improvement can be achieved [35]. From our study, irrigation at R3 was the most effective.
Drier growing conditions obviously constrained BNF. Drought in 2009 resulted in lower BNF compared with that in 2007. It has been reported that BNF is more sensitive to drought than other physiological processes [9,40] because plants have to justify the partitioning of fixed C to cope with the drought impact, and nodule formation and fixation activities are most affected [3]. Soybean has the highest BNF ability at R5 and maximum N demand occurs between stages R3 and R5 [5,7,9]. BNF is also closely related to the crop growth rate i.e., C supply [5]; less photosynthesis through the growing season in 2010 resulted in a failure in insufficient energy for BNF compared with other experimental years. From our model simulations, BNF was 41% less in 2008 compared to that in 2007 during R3-R5 because photosynthesis rates during this stage were much lower in 2008.
Applied fertiliser N is reported to inhibit BNF [14,33,41,42] because a high soil mineral N content can constrain nodule numbers and biomass [16,40]. From our simulations, these constraints became larger with delayed fertiliser application timing. This is consistent with a previous field study in which fertiliser N application at the reproductive stage gave 3.3-5.5% reduction in the contribution of BNF compared with fertiliser application at sowing [8]. A possible reason is that legume plants may suffer from N hunger for a period of 15-20 days after sowing prior to nodule development and N fixation beginning, so a small amount of fertiliser at sowing can help root formation [11] and alleviate any N shortage over this period. On the other hand, BNF is an energy intensive process and legume crops will favour mineral N whenever it is available in preference to BNF. In our study, BNF was calculated based on a root-based algorithm generated from a perennial temperate forage legume, because robust data on soybean nitrogen fixation for this region are lacking; improvements to these estimates would benefit from controlled experiments on soybean N fixation rate and the influencing factors.
However, fertilization at the beginning of the growing season (BS, BR1 and BR3) can enhance seed N content, compared with a single application during the reproductive phase, which is supported by previous studies [13,41]. In cool climate zones, starter N fertiliser can have a positive effect on early root and above-ground growth, thus enabling more soil mineral N to be taken up, which would provide more N remobilized from leaves and stems during the reproductive phase [18,43].
In order to have sustainable agricultural systems, there should ideally be a balance between N inputs and outputs, without too much N either drawn from or left over in the soil [42]. Synchronization of higher temperatures and more precipitation in July (R3-R5 phase) resulted in a higher N mineralization rate to meet the N demand of the crop. With soil total N at this site of about 8.7 Mg/ha in the top 80 cm soil depth (Table 1), the deficit (39 kg N/ha, and compensated by the soil mineralization process) under the condition of no fertiliser application implies that soil total N would be used up in around 255 years at a mean annual turnover rate of 0.39% (0.04-0.67%). The soil N budget will always be negative without external N input and would become increasingly negative with increasing seed yield, indicating a soil N deficit [14]. The suggested fertiliser N application rates from this simulation were within the range of local practices (fertiliser applied before sowing). Nitrogen losses through denitrification, leaching and surface runoff were not only determined by external N inputs and plant uptake, but also by N mineralization. It has been reported that 55% of N losses may be attributed to the release of native soil N into the environment, due to the asynchrony between soil mineralization and crop uptake in a maize-soybean rotation system in the USA [19]. In our simulations, the derived N budget included N losses through denitrification and via leaching, accounting for the impact of N mineralization and nitrification which previous studies have failed to consider due to difficulties in measurement [20,42].

Conclusions
The process-based SPACSYS model was shown to provide good simulations of the dynamics of dry matter and N content in leaves, roots and seeds, and soil water content and temperature compared with field observations from Northeast China. Scenario testing suggested that yield did not substantially increase with additional N fertiliser, with a trade-off between added fertiliser N and BNF, while irrigation at the reproductive stage improved yield in the drier years but was not necessary in the wetter years. The soil-plant N budget was negative without N fertilization. However, N losses through runoff and leaching in wetter years can account for up to 18% of input N. From the perspective of maintaining the soil N balance and achieving higher yield, applying 25-30 kg N/ha before sowing was advised for drier years and 15-20 kg N/ha for wetter years. To achieve a higher seed N content, an application rate of 55-60 kg N/ha was recommended for drier year and 45-50 kg N/ha for wetter years.

Funding:
The study was supported by the Newton Fund through the BBSRC project of China Virtual Joint Centre for Improved Nitrogen Agronomy (CINAG) (BB/N013468/1).