43 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
Table 7 Summary of the APSIM Nwheat model performance at the Eest, at PAGV, and the Bouwing in the Netherlands
Model attribute Number of paired data points
Observed range r
2 1:1a m
b RMSD
c Grain
d yield t ha−1 63
0.4–8.3 0.90
0.96 0.8
Kernels d m−2
27 12726–27916
0.55 0.95
2604 Kernel
d weight mg 27
9–37 0.79
1.04 3.9
Anthesis date day 6
14 0.76
0.98 3.7
Biomass t ha−1 129
0.03–20 0.97
1.02 1.2
LAI m m−2 126
0–5.5 0.65
1.28 1.2
Crop N kg ha−1 128
2–276 0.82
1.03 28.5
Grain d N kg ha−1
63 10–204
0.88 0.98
22.4 Grain protein
18 7.1–15.6
0.59 1.00
1.6 Soil mineral N kg ha−1
546 0–78
0.46 0.75
9 a r2 1:1=r2 for the 1 to 1 line y=x.
b Slope of linear regression forced through the origin. c Root mean square deviation.
d Including pre-maturity harvests.
tion at Swifterbant, 1974–1988 and a silty clay been summarised in Table 7. The observed grain
yields ranged from 4.5 t ha−1 or when including loam at the Bouwing, using weather data from the
nearest weather station at Wageningen, 1955–1996. pre-maturity harvests from as low as 0.4 t ha−1
to 8.3 t ha−1 Fig. 2. The yield simulations cap- tured the response to season and N-fertiliser effects
with a coefficient of determination of r2 1:1 of
3. Results
0.90 and RMSD of 0.8 t ha−1. A comparison of simulated and observed yields
Winter wheat crops were generally sown at the end of October, emerged before the beginning of
from a long-term experiment at Lovinkhoeve is shown in Fig. 3. The higher yields of this experi-
winter December, were dormant over winter, started to grow again after winter in about
ment and the reported grain yields by Spiertz and Ellen 1978 and Spiertz and Van de Haar 1978
FebruaryMarch and reached anthesis in June and maturity in August. Rainfall was usually evenly
in 1976 and 1977 for the three N treatments were well simulated. However, simulated and observed
distributed throughout the season. Soil types were favourable for root growth with a groundwater
yields showed larger differences in particular with overestimations in seasons with lower yields. Note
table at about 1 m below surface, restricting root growth in most of the years. Due to the high
that the simulated crop was killed in 1979 due to severe frost during winter by a frost kill function,
water-holding capacity of the soils and a shallow water table, periods without rainfall did not cause
which accelerates leaf senescence in the NWHEAT module. A frost-hardening routine is currently not
any severe water deficits. However, a lack of rainfall might have delayed the uptake of applied
available in NWHEAT. No simulated yields are presented for this year.
fertiliser in some years. When comparing simulated yields at the
Bouwing with regional yields from Gelderland 3.1. Performance of the APSIM Nwheat model in
the Netherlands from 1975 to 1996, much lower yields were mea-
sured than simulated during the 1970s and early 1980s, but a good correlation occurred in the last
Simulations with the APSIM Nwheat model have been compared to more than 1200 actual
simulated decade Fig. 4. The early discrepancy reflects a ‘technology gap’ between the simulations
measured values. The performance of the simula- tion model compared with field observations has
assuming disease-free crops, present-day manage-
44 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
ment and modern varieties and crop management in the seventies with partial control of diseases
e.g. mildew and in particular less N input. The yield components: number of kernels per
m 2 [r2 1:1=0.55] and specific kernel weight
r 2=0.79 were reasonably well simulated, but
were less accurate than grain yield simulations Table 7.
The coefficient of determination for the date of 50
anthesis was high r 2=0.76, Table 7. The
overall phenology simulations were found to be good.
Simulations of biomass production were very good with r
2 1:1 of 0.97 Fig. 5, Table 7. In
Fig. 3. Simulated ——— and observed : long-term winter
contrast to the biomass simulations, leaf area index
wheat experiment at Lovinkhoeve, K. B. Zwart, AB-DLO, pers. comm. grain yields with a 0 kg N ha−1, b 38 kg N ha−1,
LAI simulations were poor with a coefficient of
and c 188 kg N ha−1 from 1974 to 1987. Additional measured
determination r 2 1:1 of 0.65 Table 7 and a
grain yields with 50 and 200 kg N ha−1 in 1976 Spiertz
tendency of overpredicting LAI m=1.28, Table 7;
and Ellen, 1978 and 0 and 50 kg N ha−1 in 1977 Spiertz and
Fig. 6. Frequently, the model tended to under-
Van de Haar, 1978. No measured yields were available at every
estimate the rapid senescence of leaves when
fourth year from the long-term winter wheat experiment K.B. Zwart, pers. comm.. Note, no simulated yield was available
for 1979 due to simulated frost kill. Solid bars in c show annual rainfall amounts.
approaching maturity. At two times the East, N1 treatment, 18 July 1983 and PAGV, N1 treatment,
19 July 1983, the simulated LAI was still above 1.0 while leaves were already fully senesced in the
field experiments Fig. 6.
Total crop N kg N ha−1 simulations were reasonable with a coefficient of determination r2
1:1 of 0.82 Table 7; Fig. 7, with a tendency to
Fig. 2. Model performance for grain yields. Simulated lines and observed symbols grain yields for treatment N1 ———
and N3 6 · · · · · · at a the Eest in 1983, b PAGV in 1983
Fig. 4. Simulated grain yields with 200 kg N ha−1 at the and c the Bouwing in 1984 Table 2. d Simulated versus
observed grain yields at the Bouwing 6 ,
the Eest and Bouwing ——— and Gelderland regional yields n: Crop
Estimates, Statistics Netherlands, Voorburg from 1975 to 1996. PAGV in 1983 and the Bouwing +, the Eest
and PAGV in 1984. Small symbols are pre-maturity har- Note, no simulated yields were available for 1979, 1985 and
1987 at the Bouwing due to simulated frost kill. Solid bars show vests. The dotted line is the 1 to 1 line. Observed data after
Groot and Verberne 1991. annual rainfall amounts.
45 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
over predict slightly at lower crop N content and larger variations above 100 kg N ha−1 of crop N.
Simulated and observed patterns of total shoot biomass production, LAI and crop N for two
different N applications on a silty loam at the Eest in 1983 are shown in Fig. 8. Increasing the N
application from 0 to 160 kg N ha−1 had a large effect on observed total shoot biomass production,
which
was also
reproduced by
the model
[Fig. 8a]. The simulations appropriately reflected the differences in LAI between the two N treat-
Fig. 5. Model performance for total shoot biomass. Simulated
ments [Fig. 8b]. However, LAI was over-
versus observed total shoot biomass at the Bouwing 6 , the
predicted between the end of tillering late April
Eest and PAGV in 1983 and the Bouwing +, the
and flag leaf stage end May. The model provided
Eest and PAGV in 1984. The dotted line is the 1 to 1
a satisfactory simulation of crop N above-ground
line. Observed data after Groot and Verberne 1991.
biomass N dynamics for the two N treatments shown in Fig. 8c. Note the sudden drop of about
80 kg N ha−1 in the measured crop N amounts of the high N treatment during grain filling at 20
July, which might suggest the presence of a larger variability in the actual measurements.
The model
simulated grain
N kg ha−1
Fig. 6. Model performance for LAI. Simulated versus observed LAI at the Bouwing
6 , the Eest and PAGV in 1983
and the Bouwing +, the Eest and PAGV in 1984. The dotted line is the 1 to 1 line. Observed data after Groot
and Verberne 1991.
Fig. 8. Effects of N fertiliser on shoot biomass production, LAI Fig. 7. Model performance for total above ground crop N.
and crop N. Simulated lines and observed symbols for N1 ——— and N3 +- - - fertiliser treatments at the Eest, 1983
Simulated versus observed crop N at the Bouwing 6 , the Eest
and PAGV in 1983 and the Bouwing +, the Eest Table 2; a total shoot biomass production, b LAI and c
crop N. The arrow indicates the date of observed anthesis. and PAGV in 1984. The dotted line is the 1 to 1 line.
Observed data after Groot and Verberne 1991. Observed data after Groot and Verberne 1991.
46 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
response to N-fertiliser applications and different seasons reasonably well, resulting in a coefficient
of determination of r 2 1:1 of 0.88 Table 7;
Fig. 9. Grain protein is derived from the ratio of grain
N to grain yield. The measured protein concen- tration ranged from 7 to 16
, whereas the simu- lated protein ranged from 8 to 16
. The model was able to reproduce the general trend of grain
protein response to N-fertiliser application, but could differ in some cases by several protein
from measured values, e.g. in the Bouwing 1984 [Fig. 10a]. Grain protein
simulations had a coefficient of determination of r2 1:1 of 0.59 and
RMSD of 1.6 [ Table 7; Fig. 10b]. The dynamics of soil water contents were well
simulated for wet and dry seasons [Fig. 11a and b] when the bypass flow was neglectable and
when fluctuations of the water table were not affected by additional lateral ground water inflow,
as occasionally occurred in the Eest and PAGV in
Fig. 10. Model performance for grain protein . a Simulated
lines and observed symbols grain protein for three N treat-
ments Table 2 at PAGV in 1983 ——— and at the Bouwing in 1984
6 · · · · · · . b Simulated versus observed grain protein content at the Bouwing
6 , the Eest and
PAGV in 1983 and the Bouwing +, the Eest and PAGV in 1984. The dotted line is the 1 to 1 line. Observed
data after Groot and Verberne 1991.
1983. No such effect was apparent at the Bouwing in 1984. Observed and simulated soil water
contents in the silty clay loam at the Bouwing are shown in Fig. 11c, for three dates in the growing
season 1984. The model closely simulated soil water contents in the 1 m profile at the flag leaf
stage 31 May, beginning of grain filling 4 July and during grain filling 18 July.
It was possible to simulate the fluctuation of the groundwater table with the model by setting
km less than 1 and ks to a very low value in the bottom of the soil profile [ Table 3b]. However,
the simulated water table depths were less sensitive
Fig. 9. Model performance for grain N. Simulated lines and
than the measurements not shown due to occa-
observed symbols grain N for treatment N1 ——— and N3 D · · · · · · at a the Eest in 1983, b PAGV in 1983
sional bypass flow and additional lateral ground
and c the Bouwing in 1984 Table 2. d Simulated versus
water inflow. Nevertheless, bypass flow and addi-
observed grain N at the Bouwing 6 , the Eest
and PAGV
tional lateral ground water inflow tended not to
in 1983 and the Bouwing +, the Eest and PAGV
affect the performance of any other model compo-
in 1984. Small symbols are pre-maturity harvests. The dotted
nents, including soil N dynamics and crop growth.
line is the 1 to 1 line. Observed data after Groot and Verberne 1991.
The performance of the model in simulating
47 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
Fig. 12. Effects of N fertiliser on crop N and soil mineral N content. Simulated lines and observed symbols for a crop
N, b soil mineral N in 0–20 cm and c soil mineral N in 20– 40 cm at PAGV in 19831984. Treatment N1 ——— with
80 kg N ha−1 applied on 17 February 1984 and treatment N3
Fig. 11. Model performance for soil water. Simulated lines +- - - with 80 kg N ha−1 applied on 17 February 1984, plus
and observed symbols total soil water in 0–100 cm soil depth 120 kg N ha−1 on 14 May 1984 and 40 kg N ha−1 on 8 June
N1 treatments, Table 2 for a a silty clay loam at the Bouwing 1984. Both simulations were initialised on 17 October 1983,
in 1983 ——— and b a silty loam at PAGV in 1984 with soil N measurements from 3 November 1983. The arrows
+ · · · · · · . Solid bars show daily rainfall amounts. c indicate the dates of N applications. Observed data after Groot
Simulated and observed volumetric soil water contents on 31 and Verberne 1991.
May ———, 4 July – – – and 18 July + · · · · · · 1983 for a silty clay loam at the Bouwing 1983. Observed data after
Groot and Verberne 1991.
above-ground biomass and 27.9 kg N ha−1 in the roots root N not shown, slightly overestimating
soil mineral N was acceptable, even though only a relatively low coefficient of determination was
the measured uptake [Fig. 12a]. During the same time, 4.9 kg N ha−1 was immobilised in soil
obtained [r 2 1:1=0.46, RMSD=9; Table 7]. In
particular, the pattern of mineral N in the soil was organic matter, 1.9 kg N ha−1 was lost through
denitrification, and 0.02 kg N ha−1 was lost by generally simulated well by the model, considering
the variability and observed rapid changes. Fig. 12 NO
3 leaching not shown. These processes caused
a reduction in soil mineral N of 118 kg N ha−1 shows the dynamics of crop and soil N in two
different N treatments on a silty loam at PAGV including a monthly 3 kg N ha−1 of N deposition,
added to the surface layer on 31 May 1984, with in 19831984.
Fig. 13 shows simulated changes of N compo- most of the changes occurring in the top 20 cm
[Fig. 12b]. A further breakdown of soil organic nents within the crop–soil system. Nitrogen
dynamics after the second N application of matter pools is shown in Fig. 13b. The fresh
organic matter pool increased its N content in this 120 kg N ha−1 at 14 May 1984 are presented until
the day before the third N application at 7 June period by 5.5 kg N ha−1. This pool had received
all the initial root residues and 95 of the initial
1984, for the highest N treatment at PAGV 19831984 the same as the high N treatment in
surface residues when ploughed in the soil on 17 October 1983. The microbial biomass pool showed
Fig. 12. Fig. 13a indicates that in this period, 114.3 kg N ha−1 were stored in the above- and
an increase in N content of 2.1 kg N ha−1, whereas a N loss occurred in the slow decomposing soil
below-ground biomass, with 86.4 kg N ha−1 in the
48 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
Fig. 13. Simulated change in N components within the crop– soil system for the period after the second N application of
120 kg N ha−1 at 14 May 1984 until the day before the third N application at 7 June 1984 for the high N treatment N3 at
PAGV 1984 the same N3 treatment as in Fig. 12. a Above- and below-ground crop N - - - -, above-ground crop N – – –
–, soil organic matter N – · – · –, denitrification loss · · · · · · and soil mineral N ———. The arrow indicates a monthly N
deposition of 3 kg N ha−1, added to the surface layer on 31 May 1984. b Fresh organic matter N – · – · –, net N minerali-
sation – – – –, microbial biomass N - - - -, surface residue N · · · · · · and soil humus N ———.
Fig. 14. Simulated effects of rate and timing of N-fertiliser appli-
humus pool of 2.6 kg N ha−1 and in the surface
cation in addition to the main application in February at the
residue pool of 0.1 kg N ha −1. The loss of N is
Eest Table 6. Cumulative probability distributions of a grain
shown as negative. The net increase in soil organic
yield, b grain protein and c soil residual N for treatments A ———, B — —, C — 20 kg N ha−1 at DC39 – – – –,
matter N was due to N loss from decaying roots
D — 40 kg N ha−1 at DC39 - - - - -, E — 60 kg N ha−1 at
feeding into the fresh organic matter pool.
DC39 · · · · · · and F — 80 N kg ha−1 at DC39 – · – · –. All
However, there was a positive net mineralisation
treatments A–F
received a
February N
application.
of 4.7 kg N ha−1 in this period 14 May to 7 June.
Treatments B–F received 90 kg N ha−1 at DC23DC31.
3.2. Simulation experiments median of 8.7
. Soil residual N for treatment A, however, ranged from 9 to 20 kg N ha−1 with a
Probability distributions for grain yield, grain protein and soil residual N for six different N
median of 10 kg N ha−1. Simulated grain yields and grain protein [Fig. 14a and b] were largely
applications for the Eest based on 14 years of historical weather data from Swifterbant are pre-
affected by the second and third N application treatment B during tillering DC23 and the
sented in Fig. 14. With N applied only in February to increase the soil mineral N to 140 kg N ha−1
beginning of stem elongation DC31, respectively, but hardly affected soil residual N, except during
treatment A,
yields ranged
from 5.8
to 8.6 t ha−1, with a median =50
probability of the year 1985 with a dry period between 6 May
and 13 June average rainfall from 1974 to 1984 exceeding or 1 out of 2 years’ chance of
7.8 t ha−1 [Fig. 14a]. Grain protein for the same for this period: 76 mm. Yields for treatment B
ranged from 6.4 to 9.4 t ha−1, with a median of treatment A ranged from 7.9 to 10.5
with a
49 S. Asseng et al. European Journal of Agronomy 12 2000 37–54
8.4 t ha−1, and grain protein ranged from 11.3 to 13.5
with a median of 12.8 . Soil residual N for
treatment B ranged from 9 to 49 kg N ha−1, with only a small increase in the median, compared to
treatment A, from 10 to 11 kg N ha−1. Model simulations showed [Fig. 14a and b]
that additional N fertiliser applied at the flag leaf stage DC39 treatments C, D, E and F had
minimal effects on grain yields, but increased grain protein to medians of 13.5, 14.0, 14.9 and 15.0
when applying 20, 40, 60 and 80 kg N ha−1, respec- tively. Such an increase in applied N at DC39
hardly affected the median for residual N
13 kg N ha−1 with up to 40 kg N ha−1 fertiliser, but increased the median to 18 and 31 kg N ha−1
with 60 and 80 kg N ha−1 fertiliser applied, respec- tively [Fig. 14c]. However, the upper range in
soil residual N was affected with each additional N application at DC39, reaching 68, 87, 106 and
125 kg N ha−1 with fertiliser applications of 20, 40, 60 and 80 kg N ha−1, respectively.
In a second experiment, grain yield, grain pro- tein and soil residual N were compared for zero
0 and maximum max N applications for the Eest, based on simulations with 14 years of histori-
cal weather data from Swifterbant and for the Bouwing, based on simulations with 31 years of
Fig. 15. Simulated effects of zero 0 and maximum max N-fertiliser application at the Eest and at the Bouwing.
historical weather
data from
Wageningen
Cumulative probability distributions of a grain yield, b grain
Fig. 15. The additional N-fertiliser applications
protein and c soil residual N for the Eest 0 ——— and max
in the maximum N treatment had minor effects on
- - - - - and the Bouwing 0 – – – – and max · · · · · · .
grain yield and protein, but increased the soil residual N substantially. The highest simulated
4. Discussion