Europ. J.

Agronomy 76 (2016) 41–53

Contents lists available at ScienceDirect

European Journal of Agronomy

journal homepage: www.elsevier.com/locate/eja

Performance and sensitivity of the DSSAT crop growth model in

simulating maize yield under conservation agriculture
Marc Corbeels a,b,∗ , Guillaume Chirat a , Samir Messad c , Christian Thierfelder d
a
CIRAD, AIDA (Agro-ecology and Sustainable Intensification of Annual Crops), Av. Agropolis, 34060 Montpellier, France
b
CIMMYT, SIP (Sustainable Intensification Program), P.O. Box 1041-00621, Gigiri, Nairobi, Kenya
c
CIRAD, SELMET (Mediterranean and tropical livestock systems), Av. Agropolis, 34060 Montpellier, France
d
CIMMYT, SIP (Sustainable Intensification Program), P.O. Box MP 163, Mount Pleasant, Harare, Zimbabwe

a r t i c l e i n f o a b s t r a c t

Article history: With the practice of conservation agriculture (CA) soil water and nutrient dynamics are modified by the
Received 19 May 2015 presence of a mulch of crop residues and by reduced or no-tillage. These alterations may have impacts
Received in revised form 3 February 2016 on crop yields. The crop growth model DSSAT (Decision Support Systems for Agrotechnology Transfer)
Accepted 7 February 2016
has recently been modified and used to simulate these impacts on crop growth and yield. In this study,
Available online 17 February 2016
we applied DSSAT to a long-term experiment with maize (Zea mays L.) grown under contrasting tillage
and residue management practices in Monze, Southern Province of Zambia. The aim was (1) to assess the
Keywords:
capability of DSSAT in simulating crop responses to mulching and no-tillage, and (2) to understand the
DSSAT
Co-inertia analysis sensitivity of DSSAT model output to input parameters, with special attention to the determinants of the
Conservation agriculture model response to the practice of CA. The model was first parameterized and calibrated for the tillage
Crop growth model treatment (CP) of the experiment, and then run for the CA treatment by removing tillage and applying
Maize a mulch of crop residues in the model. In order to reproduce observed maize yields under the CP versus
Sensitivity analysis CA treatment, optimal root development in the model was restricted to the upper 22 cm soil layer in
the CP treatment, while roots could optimally develop to 100 cm depth under CA. The normalized RMSE
values between observed and simulated maize phenology and total above ground biomass and grain
yield indicated that the CA treatment was equally well simulated as the CP treatment, for which the
model was calibrated. A global sensitivity analysis using co-inertia analysis was performed to describe
the DSSAT model response to 32 model input parameters and crop management factors. Phenological
cultivar parameters were the most influential model parameters. This analysis also demonstrated that
in DSSAT mulching primarily affects the surface soil organic carbon content and secondly the total soil
moisture content, since it is negatively correlated with simulated soil water evaporation and run-off. The
correlations between the input parameters or crop management factors and the output variables were
stable over a wide range of seasonal rainfall conditions. A local sensitivity analysis of simulated maize
yield to three key parameters for the simulation of the CA practice revealed that DSSAT responds to
mulching particularly when rooting depth is restricted, i.e., when water is a critical limiting crop growth
factor. The results of this study demonstrate that DSSAT can be used to simulate crop responses to CA,
in particular through simulated mulching effects on the soil water balance, but other, often site-specific,
factors that are not modeled by DSSAT, such as plough pan formation under CP or improved soil structure
under CA, may need to be considered in the model parameterization to reproduce the observed crop yield
effects of CA versus CP.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction

Conservation agriculture (CA) is nowadays perceived as a set

of best-management practices based on no-tillage, crop residue
mulching and the use of crop rotations and/or associations, through
∗ Corresponding author at: CIMMYT, SIP (Sustainable Intensification Program), which African agriculture can combat soil degradation, secure crop
P.O. Box 1041-00621, Gigiri, Nairobi, Kenya. yields and mitigate to some extent negative effects of climate
E-mail address: [email protected] (M. Corbeels). change (Gowing and Palmer, 2008; Thierfelder et al., 2014), even

http://dx.doi.org/10.1016/j.eja.2016.02.001
1161-0301/© 2016 Elsevier B.V. All rights reserved.
42 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53

if its potential is site-specific and depends on the socio-economic able literature, or determined through model calibration. The large
environment of the farming communities (Uri, 2000; Giller et al., number of model input parameters and their uncertainty lead to
2009; Corbeels et al., 2014). Under CA management, soil water questions on how large the resulting prediction uncertainty is for
and nutrient dynamics are modified by reduced or no-tillage and different model outputs and for different plant growth situations.
the presence of a mulch of crop residues on the soil surface. To begin a DSSAT simulation, the model is informed about the
Mulching with crop residues reduces soil water evaporation and specific weather, crop and soil characteristics. The corresponding
run-off (Scopel et al., 2004), increases topsoil organic matter and input files are linked to the main structure of the model (Fig. 1),
improves near-surface soil aggregate properties (Blanco-Canqui comprising the modules for the field characterization, the initial
and Lal, 2007), with potentially positive effects on crop productivity soil conditions and the management operations. The main struc-
(Rusinamhodzi et al., 2011). ture of DSSAT is designed as a matrix of simulation treatments that
Cropping system models, such as DSSAT (Decision Support Sys- implements the selected crop and soil models to describe on a daily
tems for Agrotechnology Transfer, Jones et al., 2003) or APSIM basis the changes in plant and soil variables that occur on a specific
(Agricultural Production Systems Simulator, Keating et al., 2003), land unit (field) in response to weather and management. A detailed
are based on ecological principles for simulating crop development description of the DSSAT model with its modules is given in Jones
and growth as a function of weather conditions, soil properties and et al. (2003).
management practices (through simulated water and nutrient limi- To simulate tillage effects it is assumed in DSSAT that the fol-
tations to plant growth). This type of models has been used in recent lowing four soil properties change (Andales et al., 2000): (1) soil
years to assess and analyze the agronomic performance of CA sys- bulk density; (2) saturated soil hydraulic conductivity; (3) the soil
tems. In particular, users have compared simulated yields of a given runoff curve number, and (4) soil water content at saturation. These
crop grown under conventional tillage-based practices with those soil properties are input after a tillage event and they change back
under CA management at specific sites in diverse edaphic and cli- to a settled (user-specified) value, following an exponential curve
matic conditions (e.g., Sommer et al., 2007; MacCarthy et al., 2010; that is a function of cumulative rainfall kinetic energy since the
Gerardeaux et al., 2011; Ngwira et al., 2014). last tillage operation. Tillage events also result in a mixing of soil
DSSAT (Jones et al., 2003) is a process-based cropping system components including soil water, inorganic soil nutrients and soil
simulation model that has been regularly revised to improve the organic matter pools within the specified tillage depth. The mixing
biophysical representation of soil water, organic matter and nutri- efficiency, or percentage of soil that is mixed, is also a user-specified
ent (nitrogen and phosphorous) dynamics and their effects on crop input for each type of tillage operation. Finally, tillage increases the
growth and yield. For instance, the CENTURY soil organic mat- decomposition rate of the soil organic matter pools for a period
ter model (Parton et al., 1987) was incorporated into DSSAT by of 30 days after the tillage event (Porter et al., 2010). Without
Gijsman et al. (2002) to improve simulations of long-term soil car- tillage, all crop residues remain on the soil surface. The surface
bon and nitrogen dynamics. More recently, further modifications residues decompose over time with the occurrence of immobi-
of DSSAT were made in order to simulate the effects of tillage and lization/mineralization of nitrogen as simulated by CENTURY, and
surface crop residues on soil water and organic matter dynamics part of the organic matter transforms in more stable organic mat-
(Porter et al., 2010). DSSAT has increasingly been used as a tool ter pools in the soil (Gijsman et al., 2002). Mineralized nitrogen
to compare the performance of different cropping systems and from decomposing surface organic matter is assumed to leach into
crop production technologies (e.g., Jagtap and Abamu, 2003; Fofana the (top) soil and is available for plant uptake. Mulching effects
et al., 2005; Saseendran et al., 2007; Caviglia et al., 2013). Its over- on the soil water balance are simulated by DSSAT through three
all aim is to gain better understanding of how cropping systems soil water-related processes: (1) rainfall interception by the mulch;
and their components function and to guide decisions about trans- (2) reduction of soil evaporation rates, and (3) reduction of sur-
ferring production technologies from one location to others where face water runoff (Porter et al., 2010). DSSAT version 4.5 does not
soils and climate are different. However, the capability of DSSAT to specifically simulate effects of surface residues on soil temperature
simulate crop responses to CA practices has not yet been assessed dynamics.
thoroughly. Root growth in DSSAT is simulated as a function of above ground
In this study, we applied DSSAT to a long-term experiment com- biomass production, i.e., above ground biomass has priority for
paring crop performance of maize (Zea mays L.) under conventional assimilated carbohydrates and at the end of each day carbohydrates
tillage-based and CA practices in Monze, Southern Province, Zam- not used for above ground biomass are allocated for root growth.
bia. The objectives of the study were (1) to assess the capability of The root distribution weighing factor is used to simulate the relative
DSSAT in simulating the effect of the practice of CA (viz. no-tillage root growth in all soil layers in which roots actually occur (Ritchie,
and mulching) on crop yield, and (2) to understand the sensitivity 1998). It is multiplied by a soil water factor to obtain the actual
of DSSAT model output to input parameters, with special attention root distribution. The root distribution weighing factor is an input
to the determinants of the model response to the practice of CA. for each soil layer and reflects physical or chemical constraints on
root growth in certain soil layers. Its value ranges from 1 (indicat-
ing that the soil layer is most hospitable to root growth to near 0
2. Materials and methods indicating that the soil is inhospitable for root growth).

2.1. Overview of DSSAT 2.2. Site and field experiment

In this study we used DSSAT version 4.5, with CERES-Maize DSSAT was run using data from a field experiment with maize
as the crop model (Jones and Kiniry, 1986), CENTURY to simulate under contrasting tillage and residue management practices in
soil carbon and nitrogen dynamics (Gijsman et al., 2002), and the order to test its responses to CA practices (tillage and mulching)
Ritchie soil water balance model which uses the one dimensional and to obtain a realistic framework for model sensitivity analy-
‘tipping bucket’ approach (Ritchie et al., 2009). DSSAT is a com- sis. The field experiment was conducted by CIMMYT (International
plex non-linear dynamic model that simulates outputs such as crop Maize and Wheat Improvement Centre) at the Farmer Training
development and yield as a function of a large number of input Centre in Monze (16◦ 14 24 s, 27◦ 26 24 E, 1103 m.a.s.l.), during six
parameters, including plant and soil parameters, for which values cropping seasons from 2005 to 2011 (Thierfelder and Wall, 2009;
are commonly estimated based on field experiments or from avail- Thierfelder et al., 2013). The climate at the site is tropical wet and
 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53 43

Fig. 1. Overview of the components and modular structure of the DSSAT model.

Table 1
Values of soil parameters of the soil module of DSSAT for the maize experiment in Monze, Zambia.

Soil layer (cm) Albedo Soil Drainage Runoff

evapora- rate curve
tion limit (day−1 ) number
(mm)
surface 0.14 3 0.75 84

Clay (%) Silt (%) Organic C pH in water CEC Bulk Crop- Drained upper Saturated Hydraulic
(%) (cmol kg−1 ) density determined limit upper limit conductivity
(g cm−3 ) lower limit (cm3 cm−3 ) (cm3 cm−3 ) (cm h−1 )
(cm3 cm−3 )

0–22 16 5 0.6 4.4 3.1 1.67 0.122 0.187 0.38 2.6

23–56 37 8 0.3 5.1 4.0 1.48 0.228 0.314 0.40 0.4
57–80 38 8 0.05 5.3 5.3 1.48 0.264 0.309 0.39 0.1
81–107 44 7 0.05 5.6 5.4 1.33 0.296 0.331 0.39 0.1
>–107 43 7 0.05 5.7 6.7 1.48 0.304 0.324 0.39 0.1

dry (Aw, Köppen Climate Classification) with a unimodal rainfall the conventional tillage (using a moldboard plough) treatment (CP)
pattern (Fig. 2). Rains start in November and end in April. The with removal of the crop harvest residues, and the CA treatment
average annual rainfall at the site is 750 mm. During the dura- (CA) with the use of an animal traction direct seeder and crop
tion of the experiment, four seasons had normal rainfall, whilst residue mulching. Maize was sown between late November and
the 2006/2007 season was drier (510 mm) and that of 2007/2008 early December with a target population of 44 000 plants ha−1 . The
was wetter (1000 mm) than normal. The soil at the experimental commercial hybrid maize variety SC513 was used in 2005/2006
site is a ferric Lixisol (Thierfelder and Wall, 2009). Five soil profiles and 2006/2007, thereafter it was replaced with the variety MRI624.
were characterized on the experimental site as a variation of the Both are early to medium maturing varieties. Basal fertiliza-
same soil type. For the model simulations we considered the most tion was carried out with 165 kg ha−1 of Compound D (10:20:10,
typical soil profile for the site, as described in Table 1. N:P2 O5 :K2 O) at planting and 200 kg ha−1 urea (46% N) was applied
From ten experimental treatments that were set up in 2005, two as top-dressing in a split application, five to seven weeks after
contrasting treatments were selected for the model simulations: planting. Weed control was done by a pre-emergence application of
44 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53

Fig. 2. Average weather conditions for Monze, Zambia.

glyphosate (N-(phosphonomethyl) glycine, 41% active ingredient) timate stable soil carbon for some soil types which also contain
at a rate of 3 L ha−1 followed by regular hand-weeding as necessary significant portions of biochemically protected carbon (Six et al.,
to keep the plots free of weeds. In each cropping season, dates of 2002). The drained upper limit of plant water availability was deter-
emergence, tasseling (when 50% of the plants had mature tassels) mined through laboratory measurements; it depends solely on soil
and silking (when silks were visible outside the husks on 50% of properties. In contrast, we used a crop determined lower limit
the plants) were recorded. Dates of physiological maturity were of plant water availability defined as the lowest field-measured
available only for two seasons and were estimated for the other soil water content after plants have stopped extracting water
seasons from data provided by the seed producers. The final har- (Ogindo and Walker, 2005). Maize crop parameters are divided
vest was conducted manually by harvesting eight sub-plots of each in three subsets in the model: species-, ecotype- and cultivar-
9 m2 from each main plot. Plants were separated into cobs and veg- specific (or genetic coefficients) parameters (Fig. 1). Values used for
etative biomass and then dried for estimation of grain and straw species-specific parameters were the default values for maize in the
dry matter yields. There were four repetitions per treatment per CERES-Maize model. The values for ecotype and cultivar-specific
year. More details on the experiment and observations can be found phenological parameters (Table 2) required in CERES-Maize were
in Thierfelder and Wall (2009), Thierfelder and Wall (2010) and obtained by fitting the model to the observed dates of emergence,
Thierfelder et al. (2013). flowering, and maturity of the experimental treatment with maize
under conventional tillage in rotation with sun hemp (Crotalaria
2.3. Model settings juncea L.) and cotton (Gossypium hirsutum L.), as this treatment
was the best yielding treatment with the highest average maize
2.3.1. Model parameterisation and calibration under CP grain yields over the six seasons (6380 kg ha−1 ). The radiation use
The model was first parameterised and calibrated for the CP efficiency (RUE) was estimated by calibrating the model to fit
treatment of the experiment. It was run for the six consecutive the aboveground biomass production to observations from the CP
seasons of the experiment starting at the planting date of the treatment. The final value obtained for RUE was 3.0 g dry matter
2005/2006 season. Daily rainfall was recorded at the experimental MJ−1 PAR. Model parameters related to grain filling were obtained
site, whilst daily values for minimum and maximum temperature by fitting the model to observed grain yields in the CP treatment.
and radiation were obtained from the nearby weather station of the For the calibration we used the generalized likelihood uncer-
Monze Farm Training Centre (Fig. 2). Input parameters for the soil tainty analysis (GLUE) tool that is available in DSSAT. GLUE is a
were derived from measurements on a typical soil profile (Table Bayesian method, allowing information from different types of
1) of the experimental site. The fraction of stable carbon was ini- observations to be combined to estimate probability distributions
tialized based on silt and clay contents (see Porter et al., 2010). It of parameter values and model predictions (He et al., 2010). GLUE
represents the physically protected soil carbon, but may underes- software was run 3 times, executing 18000 tests per run. The
 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53 45

Table 2
Values of maize cultivar parameters as calibrated in the CERES-Maize crop model of DSSAT for the maize experiment in Monze, Zambia.

Cultivar P1 (◦ C day) P2 (days) P5 (◦ C days) G2 (number) G3 (mg day−1 ) PHINT (◦ C day)

SC513 240 0.0 770 550 9 55

MRI624 260 0.0 500 950 9 65

P1: Thermal time from seedlings emergence to the end of the juvenile phase (expressed in ◦ C day, above a base temperature of 8 ◦ C) during which the plant is not responsive
to changes in photoperiod. P2: Extent to which development (expressed as days) is delayed for each hour increase in photoperiod above the longest photoperiod at which
development proceeds at a maximum rate (which is considered to be 12.5 h). P5: Thermal time from silking to physiological maturity (expressed in ◦ C day above a base
temperature of 8 ◦ C). G2: Maximum possible number of kernels per plant. G3: Kernel filling rate during the linear grain filling stage and under optimum conditions (mg day−1 ).
PHINT: Phyllochron interval, i.e., the interval in thermal time (◦ C day) between successive leaf tip appearances.

following variables were fitted: dates of emergence, flowering and on the stability of the correlations between the input parameters or
maturity, grain yield and total above ground biomass at maturity. crop management factors and the output variables through model
Model performance was evaluated by calculating the normalized simulations for a drier (2006/2007) and wetter (2007/2008) season.
root mean square error (RMSE), expressed in percentage, and the Third, we determined the local sensitivity of simulated maize yield
percentage prediction deviation. Calibration was considered as to three key parameters/variables (i.e., stable soil carbon fraction,
achieved when the RMSE for all fitted maize phenology and yield depth of optimal root growth and amount of surface crop residues)
output variables was minimal. for the simulation of the CA practice (viz. no-tillage and mulching).
An important feature under the CP treatment was the exis- In the sections below we describe the details of the performed
tence of a plough pan, as suggested by the higher bulk density sensitivity analyses.
values in the 22–25 cm soil layer (unpublished data) compared
to the 0–22 cm soil layer. This plough plan restricts the pene-
tration of roots of maize plants into deeper layers, which was 2.4.1. Global sensitivity analysis
visually observed in the experiment. In contrast, in the CA treat- For the global sensitivity analysis, we used the Latin Hyper-
ment the plough plan had disappeared rapidly over time and the cube Sampling method as described by McKay et al. (1979). This
soil developed a better soil structure with cracks and voids, which is method ensures that the whole range of possible parameter val-
attributed to greater soil biological activity (e.g., more earthworms) ues is randomly sampled and that effects of interactions between
(Thierfelder et al., 2013). The importance of this phenomenon for input parameters, between input parameters and crop manage-
crop growth on similar soils in the region has already been high- ment factors, and between crop management factors on model
lighted in other studies (e.g., Materechera and Mloza-Banda, 1997). output variables are taken into account (Pathak et al., 2007).
In order to reproduce this effect of soil tillage in the model simula- The following model input parameters and management factors
tions, optimal root development rate was restricted in the model to were chosen for the sensitivity analysis (Table 3): (1) a set of input
the upper 22 cm soil layer in the CP treatment, i.e., we reduced the parameters associated with the crop cultivars, since these param-
root distribution weighing factor by 80% (from 1 to 0.2) for the soil eters are usually determined under sub-optimal plant growth
layers deeper than 22 cm resulting in restricted root growth over conditions, whilst in the model their values refer to non-limiting
depth. growth conditions; (2) a second set of model input parameters that
relate to soil moisture properties, that are often not measured in
2.3.2. Model testing for CA the field, but are inferred from laboratory measurements, which
We then ran the model for the CA treatment by turning off the may cause a laboratory-scale related systematic bias; (3) third, fac-
tillage module in DSSAT, restoring the normal root development tors related to nitrogen management and parameters of the soil
over soil depth, i.e., resetting the root distribution weighing fac- organic matter module (CENTURY) were selected, because nitro-
tor equal to 1 over the whole soil depth, and initializing DSSAT gen is considered as the main limiting nutrient for the site (and
with a mulch of crop residues with resulting modeled effects on other nutrients are not considered in the model) and, moreover
soil properties and processes. Simulated biomass and grain yield the absence of tillage under CA is assumed to affect the rate of
values were then compared with observed values and model per- nitrogen mineralization; and (4) finally, we selected the amount
formance was calculated as described for the CP treatment. The and quality (lignin content) of crop residues as key input factors
amounts of maize surface residues set in the CA simulations were for simulating the potential effect of CA on crop yield as a result
3000, 2300, 1300, 4000, 2700, 850 kg dry matter ha−1 at planting of mulching. Boundary values of the ecotype- and cultivar-specific
from 2005 to 2010. Values were estimated from the percentage input parameters were fixed to represent African maize cultivars
of observed soil cover in the experiment. The C:N ratio of maize (Table 3). Boundary values for the soil parameter values were set to
residues is a user-specified input value in DSSAT and was set at 60. characterise ferric Lixisols as described by CIMMYT (5 profiles, see
Depending on the initial amounts of residues it represents between Section 2.2). The model’s sensitivity to soil available water capac-
6 and 30 kg N ha−1 . Part of this organic nitrogen is mineralized dur- ity and soil organic carbon content was assessed for values for two
ing the growing season (simulated by the CENTURY module) and soil layers: the first layer of the top 22 cm and the second layer
becomes available as inorganic nitrogen for the crop. Values for all from 22 to 56 cm (see Table 1). Finally, the boundary values for the
other model parameters were kept the same for CP and CA. nitrogen and residue management factors were set according to the
practices adopted by the local farmers (recorded from household
2.4. Sensitivity analyses surveys) and those applied at the CIMMYT experiment in Monze. In
total, we ran 1388 combinations of input parameters and manage-
First, a global sensitivity analysis was performed to describe ment factors with R software (R Development Core Team, 2009).
the DSSAT model response to 32 model input parameters and crop The model’s sensitivity to the selected input parameters and
management factors (see Table 3). In this analysis simulations were crop management factors was assessed by looking at a set of model
conducted under ‘normal’ rainfall conditions of Monze (i.e., the output variables that are listed in Table 4. Crop grain yield and total
2009/2010 season) with the experimental field characteristics of above ground biomass are the integrative outputs of the model sim-
that season: planting date was 19th November; plant density was ulations, the other model outputs can be regarded as key variables
set at 44 000 plants ha−1 . Second, we analyzed the effects of rainfall that help in explaining simulated crop growth and yields.
46 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53

Table 3
Selected input parameters and crop management factors for the global sensitivity analysis of DSSA for the maize experiment in Monze, Zambia.

Module/Class Variable Acronym Unit Min Max Source

Genotype/ecotypes Radiation use efficiency (dry matter conversion) RUE g MJ−1 PAR 2 5 Lindquist et al. (2005)
Light extinction coefficient KCAN – 0.45 0.90 APSIM + DSSAT
◦
Thermal time from silking to effective grain filling period DSGFT C day 85 255 DSSAT (default value ± 50%)
◦
Thermal time per cm seed depth required for emergence GDDE C day cm−1 4 9 DSSAT + CIMMYT
◦
Genotype/cultivars Thermal time from emergence to end of juvenile phase P1 C day 130 380 DSSAT (range of values
◦
Thermal time from silking to physiological maturity P5 C day 600 1100 for African cultivars)+
Potential kernel number/plant G2 plant−1 400 1100 Jagtap and Abamu
Potential grain filling rate G3 mg day−1 4.0 11.5 (2003)
◦
Phyllocron interval PHINT C day 30 90
Soil description/surface Soil evaporation limit SLU1 mm 3 12 DSSAT + CIMMYT +
layer Drainage rate SLDR day−1 0.01 0.95 Gijsman et al. (2002)
Runoff curve number SLRO – 61 94
Soil description/first Lower limit SLLL.22 cm3 cm−3 0.02 0.25
soil layer (0–22 cm) Drained upper limit SDUL.22 cm3 cm−3 0.08 0.45
Saturated upper limit SSAT.22 cm3 cm−3 0.3 0.6
Bulk density SADM.22 g cm−3 0.8 1.8
Total organic C SAOC.22 % 0.2 3.0
Stable organic C SASC.22 % 60 90
Soil description/second Lower limit SLLL.56 cm3 cm−3 0.02 0.25
soil layer (22–56 cm) Drained upper limit SDUL.56 cm3 cm−3 0.08 0.45
Saturated upper limit SSAT.56 cm3 cm−3 0.3 0.6
Bulk density SADM.56 g cm−3 0.8 1.8
Total organic C SAOC.56 % 0.1 1.0
Stable organic C SASC.56 % 60 90
Main structure/initial Initial soil volumetric water content SH2O cm3 cm−3 0.0 0.3 CIMMYT
conditions Initial soil nitrate content SNO3 g N Mg−1 soil 0 10
Main N at seeding (0 DAP* ) FAMN.0 kg N ha−1 0 20 CIMMYT
structure/inorganic N at 30 DAP FAMN.30 kg N ha−1 0 50
fertilizers N at 50 DAP FAMN.50 kg N ha−1 0 50
Main structure/organic Amount of crop residues (dry matter) at planting RAMT kg ha−1 0 6000 CIMMYT + Waddington
fertilizers Lignin content crop residues PSLIG (fraction) 0.05 0.20 and Karigwindi (2004)
N content crop residues SCN % 0.5 2.0
*
DAP: days after planting.

Table 4
Selected output parameters for the global sensitivity analysis of DSSAT for the maize experiment in Monze, Zambia.

Category Variable Unit Acronym

Crop growth and Grain yield kg ha−1 Yield

development Vegetative above ground biomass (straw yield) kg ha−1 Biom
LAI max – LaiM
Cumulative plant N uptake kg N ha−1 Nup
Emergence Days after planting Emer
Silking Days after planting Silk
Maturity Days after planting Mat
Cumulative plant transpiration mm Transpi
Soil water Cumulative soil evaporation mm sEvap
Cumulative runoff mm Run
Total soil moisture at maturity mm sMoist
Soil fertility Cumulative net N mineralization kg N ha−1 netNminer
Total soil N kg N ha−1 sN
Surface organic C kg C ha−1 surfC

To summarize, the sensitivity analysis design comprises two output variables and relationships between input parameters or
tables: (1) an input table with 1388 lines that correspond to the factors and output variables. It can deal with a large number of
number of simulations (one simulation for one combination of val- input parameters or factors and with co-linearity between output
ues for input parameters or management factors), and 32 columns variables, and can take account of an imbalance between the num-
corresponding to the selected parameters and management factors ber of simulations and the number of inputs, in contrast to other
(see Table 3) and (2) an output table, with the 1388 lines (number of data coupling methods known as principal component analysis and
simulations) and 14 columns that are the 14 variables of the model canonical correlation analysis (Thioulouse and Lobry, 1995). Theory
response (see Table 4). and details of the algorithms of co-inertia analysis are described in
The method used for coupling the two tables was co-inertia Chessel and Mercier (1993) and Dolédec and Chessel (1994).
analysis, which can be considered as an alternative to the clas-
sical multivariate methods based on variance decomposition. It 2.4.2. Weather effects on the model input/output relationships
provides an overview of the linear relationships between the input The effect of weather conditions (in particular rainfall) on the
and the output tables and produces scores that are the result of the stability of the relationships between input parameters or factors
maximization of both the covariance of parameters or variables and output variables was analyzed by repeating the 1388 sim-
belonging to the same table and the covariance of parameters or ulations performed under the climatic conditions of 2009/2010
variables from one table to another. In other words, co-inertia anal- for a drier (2006/2007) and a wetter (2007/2008) cropping sea-
ysis helps to synthesize at the same time the redundancy between son. The stability of the model response was assessed using the
 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53 47

RV-correlation coefficient, a multivariate generalization of the model outputs of maize phenology, total above ground biomass and
squared Pearson’s correlation coefficient (Robert and Escoufier, grain yield compared to observed values are shown in Fig. 4.
1976). Thus, the RV-coefficient (0 = not correlated, 1 = correlated) For the variety SC513, the normalized RMSEs for grain yield
measures the reproducibility of the correlation between the model and phenology predictions were 12% and less than 6%, respectively
input and the output tables from one season to another. (Table 5). For the variety MRI624, the normalized RMSEs indicate
that the CA treatment was simulated equally well as the CP treat-
2.4.3. Model sensitivity to CA-related parameters ment: a normalized RMSE of about 10% for the phenology and of
A local sensitivity analysis was performed in order to better about 25% for grain yield (Table 5). Percentage prediction devia-
understand the potential effects of the CA practice (no-tillage and tions for total above ground biomass ranged between −14% and
mulching) on crop yield. The model’s sensitivity to the following +8% (Fig. 4B). The box plots (Figs. 3 B and 4 B) also illustrate the
three model parameters and variables was analyzed: the stable large variability in observed maize above ground biomass and grain
(or passive—see CENTURY, Parton et al., 1987) soil carbon fraction yield for a given season, a feature which was not captured by the
(fixed at 60%, 70%, 80% and 90% of the total organic carbon in the deterministic DSSAT model.
soil), the depth of optimal root growth (limited at 22, 30, 56 and
100 cm) and the amount of crop residues left on the soil surface at 3.2. Global sensitivity analysis of model output
planting (0, 1270, 2700 and 3940 kg dry matter ha−1 ). The combi-
nation of parameter values that represents best the conventional Relationships between the model input parame-
ploughing treatment is the stable soil carbon set at the lowest value, ters/management factors and the model output variables were
i.e., 60% (because of the tillage effects on soil aggregation and pro- first explored for a season with normal rainfall (2009/2010). Fig. 5
tected soil carbon), a depth of optimal root development restricted shows the results of the co-inertia analysis, linking the tables of
to 22 cm (because of the plough pan between 20 and 25 cm) and no model input and output variables. Four principal components
mulch of maize residues. The combination of values representing were a priori pertinent (see the Eigenvalues, Fig. 5) for describing
the CA treatment is the stable soil carbon set at 90%, an optimal root the relationships. However, since the fourth axis did not increase
growth over 100 cm soil depth and a mulch amount set at 1270 kg the global covariance, we limit our discussion to the first three
dry matter ha−1 or more. principal axes. Each principal axis is associated with an eigenvalue
and represents a portion of the explained covariance. Parameters
that have the highest scores are represented by the longest arrows.
3. Results
Large arrows that point in the same direction have a strong positive
correlation, whilst large arrows that point in opposite directions
3.1. Zambian case study: model calibration and testing
have a strong negative correlation, and perpendicular arrows are
not correlated. Input parameters contributing less to the model
Table 1 and 2 show the combination of values for selected model
response (the gray zones) are zoomed for better visualization.
parameters that ensured the best fit to the observed data of the
Acronyms used for input parameters/factors are those from DSSAT,
CP treatment, while Table 5 gives an overview of the capability of
explained in Table 3. Output variables and their abbreviations are
DSSAT to reproduce observed maize development and yield data
described in Table 4.
from 2005 to 2011 under the CP and CA treatment, respectively.
On the two first principal components (Fig. 5, top, right),
four model input parameters/factors were particularly well rep-
3.1.1. Model calibration—CP treatment resented: thermal time from emergence to end of juvenile phase
Results of the model calibration for the CP treatment during the (P1), thermal time from silking to physiological maturity (P5) and
six seasons of the experiment are shown in Fig. 3. We chose not to the phyllocron or interval time between appearances of successive
consider the seasons 2005/2006 and 2008/2009 for calibration of leaves (PHINT), and the amount of mulch (RAMT), meaning that
the varieties SC513 and MRI624, respectively, since we were not they contributed highly to the model response. The large arrow size
able to fit the model well enough to the observed yield data of of the amount of mulch (RAMT) in the factorial—axis 1, axis 2—plane
these seasons (Fig. 3, Table 5), probably because other factors than was partly related to the large variation in values of RAMT (from
those simulated by DSSAT may have had a substantial effect on crop 0 to 6 Mg ha−1 ). The first axis was mainly defined by P1 that was
growth during these seasons. In 2005/2006, the model largely over- positively and highly correlated to two groups of output variables,
predicted (+51% prediction deviation) total above ground biomass, including respectively above ground biomass (Biom) and maturity
resulting also in an overestimation (+40% prediction deviation) of date (Mat). This means that P1 contributed largely to the variabil-
grain yield. In 2008/2009, grain yield was overestimated (+42% pre- ity of these crop production variables. P5 was positively and highly
diction deviation), while above ground biomass production was correlated to P1. From the arrow size, we can see that the model
reasonable well predicted (+10% prediction deviation). The per- response was less sensitive to P5 than to P1. The factorial—axis
centage prediction deviation for the harvest index during that 1, axis 3—plane (Fig. 5, bottom, left) shows that P5 affected prin-
season was +28%, which suggests that the partitioning of dry mat- cipally the maturity date (Mat), while the silking date (Silk) was
ter to grains was not correctly reproduced. For the other seasons, more specifically affected by P1. The co-inertia analysis also shows
the model predicted total above ground biomass with prediction that PHINT was negatively correlated with total above ground
deviations between 0% and 4% (Fig. 3B). Normalized RMSE values biomass (Biom), leaf area index (LAIm) and cumulative crop tran-
for the grain yield simulations with model calibration were 10% for spiration (Transpi), while a positive correlation existed between
SC513 and 26% for MRI624 (Table 5). PHINT and cumulative soil evaporation (sEvap) in the factorial—axis
1, axis 2—plane, and between PHINT and total soil moisture content
3.1.2. Model testing—CA treatment (sMoist) in the factorial—axis 1, axis 3—plane. Indeed, the DSSAT
The above CP simulations were performed with optimal root model mechanisms cause that a delay in leaf development results
development limited to the first 22 cm, reproducing the effects of a in less transpiration by the crop, more soil moisture and eventually
plough pan. For the model simulations of the CA treatment, optimal more soil evaporation. From the arrow sizes and locations of the
root development was no longer restricted to the top 22 cm soil amount of mulch (RAMT), the soil drainage rate (SLDR) and of the
depth but extended to 100 cm, tillage was removed and a mulch variables sEvap and surfC, we can infer that the second axis was
of maize residues was added at planting. The resulting simulated mainly defined by the soil surface properties, while the third axis
48 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53

Table 5
Normalized root mean square errors (RMSE) between observed and simulated values for the DSSAT runs for the maize experiment in Monze, Zambia. Season 1 corresponds
to the 2005–2006 growing season, season 6 to 2010–2011. The seasons 1 (2005/2006) and 4 (2008/2009) were not considered for calibration of the varieties SC513 and
MRI624, respectively, in the CP treatment.

Treatment Variable Variety SC513 Variety MRI624

Seasons 1 and 2 Season 2 (calibration) Seasons 3 to 6 Seasons 3, 5 and 6 (calibration)

CP Silking date 3.0% 4.0% 7.2% 7.1%

Maturity date 2.4% 3.4% 9.3% 8.6%
Grain yield 25% 10% 31% 26%
Vegetative above ground biomass 38% 10% 23% 24%
CA Silking date 5.3% 10%
Maturity date 2.4% 11%
Grain yield 12% 23%
Vegetative above ground biomass 13% 26%

CP: conventional tillage treatment. CA: conservation agriculture treatment.

(A) (B)

Fig. 3. (A) Comparison of observed and simulated (DSSAT model) maize phenology and grain yield, and (B) observed (box-and-whisker plots) and simulated (points, DSSAT
model) above ground biomass for the CP treatment of the experiment at Monze, Zambia (2005/2006–2010/11 growing season). Model predictions were compared to the
mean values of observations and expressed as percentage prediction deviation.

was mainly defined by the properties of the topsoil layer (SAOC.22, The genetic coefficients, the mineral nitrogen fertilization and the
SASC.22 and SDUL.22, see Table 3). In both factorial planes (Fig. 5), organic carbon of the first soil layer were the most determining
RAMT affected primarily and positively the surface organic carbon factors for the simulated crop production variables. The amount of
content (surfC) and to a lesser extent the total soil moisture content mulch was explaining variability in grain yield to a relatively small
(sMoist), as the latter was negatively correlated with cumulative extent. This is consistent with the fact that in DSSAT crop residue
soil evaporation (sEvap) and cumulative water run-off (Run). mulching has no direct effect on simulated crop yield. The appli-
Three main groups of crop production output variables (Table 4) cation of mulch had mainly an effect on simulated organic carbon
could be distinguished on the first two principal components (Fig. 5, content, (in the first place in the surface soil layer, but also in the
top, left). A first group was associated with the first axis and includes soil layer below), and on simulated soil evaporation.
total above ground biomass (Biom), maximum LAI (LaiM), cumu-
lative crop transpiration (Transpi) and cumulative crop nitrogen
3.3. Effect of seasonal conditions on the stability of the model
uptake (Nup). A second group that was projected on the axes 1
response
and 2, included the silking (Silk) and the maturity (Mat) dates. A
third group was composed of grain yield (Yield) and was positively
To compare the possible impact of a drier (2006/2007,
correlated with the first axis and thus the first group of output vari-
510 mm rainfall) or wetter cropping season (2007/2008, 1000 mm
ables, but the size of the arrow indicated that its variance was more
rainfall) on the relationships between the model input param-
diffusely represented than those of the other groups. This means
eters/management factors and the output variables, RV values
that grain yield was not strongly determined by a single model
were calculated. These RV values indicated that the relationships
input parameter or factor. The sizes of the arrows of the input
were very stable: RV = 0.98 between 2009/2010 and 2006/2007,
parameters/factors were altogether relatively small (Fig. 5, right).
RV = 0.98 between 2009/2010 and 2007/2008, and RV = 0.97
 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53 49

(A) (B)

Fig. 4. (A) Comparison of observed and simulated (DSSAT model) maize phenology and grain yield, and (B) observed (box-and-whisker plots) and simulated (points, DSSAT
model) aboveground biomass for the CA treatment of the experiment at Monze, Zambia (2005/2006–2010/11 growing season). Model predictions were compared to the
mean values of observations and expressed as percentage prediction deviation.

between 2006/2007 and 2007/2008. Therefore, the findings on rate is simulated through an empirical non-linear function that
model behavior described in the previous section were valid for determines that the deeper a soil layer, the lesser the absolute root
a relatively wide range of crop growing conditions. growth in that layer is. Furthermore, the simulations showed that
the presence of a mulch of crop residues had a strong effect on grain
3.4. Local model sensitivity to three key model parameters for CA yield, when root growth was constrained, but the effect was small
model simulations with deep optimal rooting and, thus, water uptake from deeper soil
layers.
We analyzed the sensitivity of the model response to three This analysis illustrated the interactive effects of these model
model parameters/factors that are key for simulating the CA prac- parameters/factors on predicted grain yield. DSSAT responded to
tice, i.e., the soil depth for optimal root development, the amount crop residue mulching particularly if rooting depth is restricted e.g.,
of mulch and the stable soil carbon fraction. As expected, highest by a hard plough pan.
simulated maize grain yield occurred with the maximal depth for
optimal root development (100 cm depth), the highest amount of
mulch (3940 kg ha−1 ), and the lowest stable carbon fraction (60% of 4. Discussion
the total soil carbon), which results in the relatively highest min-
eralization of soil organic matter. Similarly, lowest grain yield was 4.1. Need for a global sensitivity analysis
simulated with the minimal depth for optimal root development
(22 cm depth), the highest fraction of stable soil carbon (90% of the The initial development of DSSAT began in the late eighties with
total soil carbon), and without mulching. However, within these the aim to integrate knowledge about crops, soils, climate and
boundaries, the simulated effects of the stable soil carbon fraction management for making better decisions about transferring crop
(Fig. 6A), the depth for optimal root growth (Fig. 6B) and the mulch production technologies from one location to others where soils
amount (Fig. 6C) were not linear. For instance, in the case of a 22 cm and climate differ (Jones et al., 2003). Before, researchers in crop
rooting depth without mulching (Fig. 6A), a 10% reduction of the genetics, plant physiology and soil science worked from their own
stable soil carbon led to a grain yield increase of 0.1, 1.4 and 0.4% disciplines to advance on the development of specific components
for initial stable carbon contents of, respectively, 90%, 80% and 70% of crop and soil models. These efforts resulted in the development of
of the total soil carbon. In the case of a 100 cm rooting depth and e.g., the CERES-Maize model (Jones and Kiniry, 1986), and the CEN-
of 3940 kg mulch ha−1 , the yield increase became 0.5, 0.2 and 0.1%. TURY soil model (Parton et al., 1987). With the development of the
These results show that simulated grain yield was weakly affected cropping system model DSSAT, these earlier models were revised
by changes in the stable soil carbon fraction in the context of this to make them compatible and additional crop and soil models were
study (Lixisol, weather conditions of 2009/2010, >100 kg N ha−1 built (Jones et al., 2003). DSSAT kept evolving during more recent
of chemical fertilizer application). On the other hand, an increase years, with e.g. the incorporation of the surface crop residue mod-
of the depth for optimal root growth from 22 to 30 cm led to a ule (Porter et al., 2010). The increasing model complexity led to an
yield gain of between 3.3 and 8.7%. A deeper optimal root devel- increasing difficulty on how to handle a tool that suffers from ‘over-
opment permits a better soil water use. Interestingly, the yield parameterization’ in the sense that given model outputs could be
increase remained weak when rooting depth increased from 30 to obtained from different combinations of (collinear) parameter val-
56 cm. Herewith, we should note that in DSSAT, the root growth ues. In other words, the effect of one parameter on model output
50 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53

Output var. scores Input param. scores

3.5
Eigenvalues
On axis 1 and axis 2:

3.0
(axes)

2.5
Axis 2 d = 0.2 Axis 2
RAMT

2.0
1.5
surfC

1.0
0.5
Biom
LaiM

0.0
SLDR
sMoist KCAN
Nup
Transpi SAOC.56
FAMN.50 RUE
FAMN.30 SADM.56
SASC.22 SDUL.56
Yield SAOC.22 DSGFT
SLLL.56 SNO3 SASC.56
netNminer
Axis 1 Axis 1
SLLL.22

PSLIG
sN
Emer
P1 GDDE
SADM.22

SLRO
SSAT.56

Run P5
SDUL.22
Silk
SSAT.22
Mat
G2
SH2O
G3
FAMN.0

sEvap

SLU1
SCN
PHINT

On axis 1 and axis 3:

Axis 3 d = 0.2 Axis 3

netNminer

SAOC.22

sMoist
SASC.22
Nup

Yield
PHINT
sN SADM.56 G3
SAOC.56 SDUL.22

SASC.56 G2
P5 SCN
sEvap SDUL.56
SLLL.56
SLLL.22
RAMT
Mat surfC FAMN.50 SSAT.22 RUE

Axis 1 Run Axis 1 FAMN.30

SH2O
SNO3
FAMN.0
Silk
GDDE
LaiM Emer PSLIG
Biom SLDR
P1 SADM.22
DSGFT
Transpi SSAT.56
SLRO KCAN

SLU1

Fig. 5. Co-inertia analysis plots of the results of the global sensitivity analysis of the DSSAT model: top, factorial—axis 1, axis 2—plane for model output variables (left) and
input parameters (right), and bottom, factorial—axis 1, axis 3—plane for model output variables (left) and input parameters (right). DSSAT was run for a maize experiment in
Monze, Zambia. Eigenvalues (total covariance explained by each co-inertia axis) are shown in the inset (top right). Input parameters contributing less to the model response
(the grey zones) are zoomed for better visualization. Acronyms used for input parameters are those from DSSAT, explained in Table 3. Acronyms for output variables are
given in Table 4. See text for further explanation.

is not only correlated with the model structure, but also with the variance and, thus, do not require an accurate determination. The
values of other parameters and input data (Reichert and Omlin, sensitivity analysis evaluated the contribution of different model
1997; Wallach et al., 2002). This makes model parameterization parameters to the model output.
and use difficult. Sensitivity analysis is a useful means for identi-
fying the important parameters that govern model output and, in
4.2. The co-inertia method
consequence, whose values need to be quantified carefully.
With the parametrization of DSSAT for a maize experiment in
There is no unique approach for sensitivity analysis of sim-
Zambia under contrasting tillage and residue management prac-
ulation models. Overall, two major categories are distinguished
tices, we faced difficulties in accurately quantifying values of
(Cariboni et al., 2007): local and global sensitivity analysis. The local
various crop parameters such as the genetic cultivar coefficients
sensitivity analysis examines the local response of model outputs
or the radiation use efficiency, and of soil parameters such as soil
by varying values of input parameters one at a time while hold-
water evaporation limit, drainage rate or stable organic carbon.
ing the values of other parameters fixed. Local sensitivity analyses
Even with available field data on soil properties, crop phenology,
are relatively easily implemented and have a low computational
biomass and grain yield, major assumptions were necessary to set
cost, but the results depend to a great extent on the initial value of
the values of several model parameters. On the other hand, some
the input parameters. Instead, global sensitivity analysis explores
of these parameters may account only for a small part of the output
the entire multi-dimensional parameter space simultaneously, i.e.,
 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53 51

Fig. 6. Results of the local sensitivity analysis of DSSAT showing maize grain yield responses as function of the stable soil carbon fraction (A), the depth of optimal root
development (B) and the amount of crop residue mulch at planting (C) for the maize experiment at Monze, Zambia.

for a specific output variable the influence of single parameters while under semi-arid climates and on poor soils the contribution
and the interactions between parameters are quantified. Several of genetic cultivar input parameters was largely reduced at the
methods exist for global sensitivity analysis, based on variance expense of model parameters describing soil water and nutrient
decomposition (e.g., the method of Sobol’), on regressions or on dynamics.
screening (e.g., the Morris method). Confalonieri et al. (2010) com-
pared different global sensitivity analysis approaches for a rice
growth model (WARM, Water Accounting Rice Model) and con- 4.3. Parameterisation of DSSAT
cluded that all resulted in relative similar outcomes. The simplest
among the methods used (i.e., the Morris method) produced results Our parameterization of the genetic cultivar coefficients
comparable to those obtained by methods more computationally (Table 2) and soil parameters (Table 1) in DSSAT led to a reasonably
expensive. We argue that besides the handling time, visualization good reproduction of the observed maize development, biomass
of the results is an important criterion for the choice of a sensi- and grain yield values over the six cropping seasons both for the
tivity analysis method, as it facilitates the interpretation of high CP and CA treatments, as indicated by the RMSE values between
dimensional data. observed and simulated data. Phenology was well predicted for
Co-inertia is a multivariate method that allows the analysis of all seasons (Figs. 3 A and 4 A). Above ground biomass and grain
a pair of numerical data tables in a robust way, i.e., regardless of yields were, however, not well reproduced for all six cropping sea-
the number of columns or lines, and with possible co-linearity sons (Figs. 3 and 4). In particular, biomass and grain yields of the
between the variables. We used this method for our global sen- CP treatment during the first season were strongly overestimated.
sitivity analysis. The co-inertia analysis searches for pairs of axes The reasons for this were not clear. Surprisingly, biomass and grain
with maximum covariance (Dolédec and Chessel, 1994). It adapts yields of the CA treatment during this season were well fitted by
a multi-dimensional cloud of data points in such a way that when the model. Similarly, grain yield of the CP treatment for the sea-
it is projected onto a two dimensional space any intrinsic patterns son 2008/2009 was not well reproduced and overestimated by the
the data may possess becomes apparent upon visual inspection. model. Here, it has to be noted that the observed harvest index was
The ordination diagrams or factorial planes are thus an important unusually low (0.38), while the simulated harvest index was rather
aid in the interpretation of the data and helped us to understand high (0.52).
the overall functioning of the DSSAT model by showing patterns of The cultivar parameters, ‘P1 , thermal time from seedlings emer-
co-variability between the model parameters and between model gence to the end of the juvenile phase, ‘P5 , the thermal time from
parameters and output variables (Fig. 5). The genetic cultivar coef- silking to maturity and the end of the juvenile phase and ‘PHINT’,
ficients P1 (thermal time from emergence to end of juvenile phase), phyllochron interval, i.e., the interval in thermal time between
P5 (thermal time from silking to physiological maturity) and PHINT successive leaf tip appearances are most influential model input
(phyllocron interval) were found to determine largely the crop parameters for which accurate values have to be determined. End of
production variables, including grain yield. Our results were in juvenile phase is defined as the time at which tassel initiation can be
agreement with those of Jones et al. (2012) who explored the DSSAT observed on 50% of the observed plants, and should be determined
model response in terms of grain yield to 17 crop and soil param- through destructive sampling by dissecting plants and observing
eters using the Sobol’s method. These authors concluded that in the apical meristem using a microscope for any development of
non-limiting plant growth environments simulated grain yields floral buds at the 2–3 days interval starting from the 10th day after
are principally determined by the genetic cultivar coefficients, emergence (Gungula et al., 2003). PHINT is estimated by the inverse
of the slope of the linear regression of the number of accumulated
52 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53

or emerged leaves on the main stem against accumulated thermal and observations. On the other hand, those parameters with a small
time. Determining the number of fully expanded leaves and leaf tips contribution to model output can be excluded from the calibration
needs accurate (and weekly) monitoring of a sufficient number of exercise and can be set equal to any value within their range. This
plants under field conditions. For P5, time of silking is recorded in contributes to a simplification of model use and is useful for calibra-
the field when silks are noticed on 50% of the observed plants. In a tion of this type of complex crop growth models for multiple sites
similar way, time of physiological maturity is recorded in the field or at a regional scale. Under the conditions of our study, genetic
as the day when 50% of the grains in each observed ear have formed cultivar coefficients were the most influential model parameters.
a black layer, indicating that no further accumulation of assimilates
is possible. Acknowledgments

4.4. DSSAT response to CA practice This study has been carried out as part of the CA2Africa
CSA-SA project (no. 245347), EU 7th Framework Programme: ‘Con-
The co-inertia analysis scores (Fig. 5) illustrated that the applied servation Agriculture in AFRICA: Analyzing and FoReseeing its
amount of mulch (RAMT) strongly and positively affects simulated Impact—Comprehending its Adoption’. The authors are grateful to
soil surface carbon content (surfC) and soil moisture (sMoist), and the colleagues of the AIDA research unit (CIRAD) for the discussions
negatively soil evaporation (sEvap) and runoff (Run). Consequently, on the use of cropping systems models.
under drought stress the practice of mulching affects simulated
grain yield positively. Results from the local sensitivity analysis cor- References
roborated this. The application of a mulch of crop residues affected
more positively simulated grain yield when water uptake was lim- Adeoye, K.B., Mohamed-Saleem, M.A., 1990. Comparison of effects of some tillage
methods on soil physical properties and yield of maize and stylo in a degraded
ited as a result of limited optimal root growth, as compared to a ferruginous tropical soil. Soil Tillage Res. 18, 63–72.
situation with deep optimal root development and greater access Andales, A.A., Batchelor, W.D., Anderson, C.E., Farnham, D.E., Whigham, D.K., 2000.
to more soil water. Incorporating tillage effects into a soybean model. Agric. Syst. 66, 69–98.
Blanco-Canqui, H., Lal, R., 2007. Soil structure and organic carbon relationships
To be able to reproduce with DSSAT the yield increase under CA following 10 years of wheat straw management in no-till. Soil Tillage Res. 95,
compared to CP, it was necessary to include a rooting depth effect, 240–254.
in addition to the mulch and the no-tillage effects. Under CP opti- Cariboni, J., Gatelli, D., Liska, R., Saltelli, A., 2007. The role of sensitivity analysis in
ecological modelling. Ecol. Mode 203, 167–182.
mal root development was restricted to the plough layer (22 cm),
Caviglia, O.P., Sadras, V.O., Andrade, F.H., 2013. Modelling long-term effects of
while under CA roots could optimally develop to 100 cm depth. The cropping intensification reveals increased water and radiation productivity in
occurrence of a hard pan caused by repeated ploughing to the same the South-eastern Pampas. Field Crops Res. 149, 300–311.
depth is a common feature in loam and clay soils (Materechera and Chessel, D., Mercier, P., 1993. Couplage de triplets statistiques et liaisons
espèces-environnement. In: Lebreton, J.D., Asselain, B. (Eds.), Biométrie et
Mloza-Banda, 1997). A hard pan restricts root growth and water Environnement. Masson, Paris (France), pp. 15–44 (French).
percolation, which can have severe effects on plant growth, espe- Confalonieri, R., Bellocchi, G., Bregaglio, S., Donatelli, M., Acutis, M., 2010.
cially if the topsoil dries out (Adeoye and Mohamed-Saleem, 1990). Comparison of sensitivity analysis techniques: a case study with the rice
model WARM. Ecol. Model. 221, 1897–1906.
Finally, in our study, simulated effects of crop residue mulching Corbeels, M., De Graaff, J., Ndah, H.T., Penot, E., Baudron, F., Naudin, K., Andrieu, N.,
on soil carbon content and nitrogen immobilization/mineralization Chirat, G., Schuler, J., Nyagumbo, I., Rusinamhodzi, L., Traore, K., Mzoba, H.D.,
were relatively small and had no effect on grain yield given the Adolwa, I.S., 2014. Understanding the impact and adoption of conservation
agriculture in Africa: a multi-scale analysis. Agric. Ecosyst. Environ. 187,
mineral fertilizer application of 108 kg N ha−1 year−1 . (Simulated) 155–170.
effects of the CA versus CP treatment on soil carbon become pro- Dolédec, S., Chessel, D., 1994. Co-inertia analysis: an alternative method for
nounced in the long term (Corbeels et al., 2014). studying species-environment relationships. Freshw. Biol. 31, 277–294.
Fofana, B., Tamélokpo, A., Wopereis, M.C.S., Breman, H., Dzotsi, K., Carsky, R.J.,
2005. Nitrogen use efficiency by maize as affected by a mucuna short fallow
and P application in the coastal savanna of West Africa. Nutr. Cycl. Agroecosyst.
5. Conclusions
71, 227–237.
Gerardeaux, E., Giner, M., Ramanantsoanirina, A., Dusserre, J., 2011. Positive effects
The results of our study illustrate that in order to simulate yield of climate change on rice in Madagascar. Agron. Sustain. Dev., 1–9.
Gijsman, A.J., Hoogenboom, G., Parton, W.J., Kerridge, P.C., 2002. Modifying DSSAT
effects of CA, an agronomic diagnosis is required on what the site-
crop models for low-input agricultural systems using a soil organic
specific factors are that explain the yield differences between CA matter-residue module from CENTURY. Agron. J. 94, 462–474.
and CP. Some of these factors are not mechanistically simulated Giller, K.E., Witter, E., Corbeels, M., Tittonell, P., 2009. Conservation agriculture and
by DSSAT, but can be incorporated in the model through a proper smallholder farming in Africa: the heretics’ view. Field Crops Res. 114, 23–34.
Gowing, J.W., Palmer, M., 2008. Sustainable agricultural development in
parameterisation of relevant parameters. This is the case for e.g., sub-Saharan Africa: the case for paradigm shift in land husbandry. Soil Use
soil structure differences as a result of the practice of CA versus Manage. 24, 92–99.
CP, which may have an effect on crop yield. DSSAT does not simu- Gungula, D.T., Kling, J.G., Togun, A.O., 2003. CERES-maize predictions of maize
phenology under nitrogen-stressed conditions in Nigeria. Agron. J. 95,
late these soil structural effects. In our study we had to introduce 892–899.
a rooting depth effect in order to reproduce the observed maize He, J., Jones, J.W., Graham, W.D., Dukes, M.D., 2010. Influence of likelihood function
yield differences between the CP (tillage and no residues) and CA choice for estimating crop model parameters using the generalized likelihood
uncertainty estimation method. Agric. Syst. 103, 256–264.
(no-tillage and crop residue mulching). Under CP the optimal root Jagtap, S.S., Abamu, F.J., 2003. Matching improved maize production technologies
growth was restricted to the upper 22 cm soil layer as a result of to the resource base of farmers in a moist savanna. Agric. Syst. 76, 1067–1084.
the formation of a hard pan just below ploughing depth, whilst Jones, J.W., Hoogenboom, G., Porter, C.H., Boote, K.J., Batchelor, W.D., Hunt, L.A.,
Wilkens, P.W., Singh, U., Gijsman, A.J., Ritchie, J.T., 2003. The DSSAT cropping
under CA optimal root growth in the model was unrestricted. Sim-
system model. Eur. J. Agron. 18, 235–265.
ulated effects of mulching on yields were more pronounced when Jones, J.W., Kiniry, J.R. (Eds.), 1986. CERES-maize, a simulation model of maize
the depth for optimal root development, and thus crop water (and growth and development. T&M University Press, College Station.
Jones, J.W., Naab, J., Fatondji, D., Dzotsi, K.A., Adiku, S., He, J., 2012. Uncertainties in
nutrient) uptake was restricted.
simulating crop performance in degraded soils and low input production
Co-inertia analysis was used in order to analyze and visualize systems. In: Kihara, J., Fatondji, D., Jones, J.W., Hoogenboom, G., Tabo, R.,
how 16 DSSAT model output variables respond to 32 input param- Bationo, A. (Eds.), Improving Soil Fertility Recommendations in Africa Using
eters, considering parameter correlations and nonlinear relations. Decision Support for Agro-technology Transfers (DSSAT). Springer, pp. 43–59.
Keating, B.A., Carberry, P.S., Hammer, G.L., Probert, M.E., Robertson, M.J.,
Co-inertia analysis can be used to visually inspect and identify influ- Holzworth, D., Huth, N.I., Hargreaves, J.N.G., Meinke, H., Hochman, Z., McLean,
ential model parameters that must be estimated from experiments G., Verburg, K., Snow, V., Dimes, J.P., Silburn, M., Wang, E., Brown, S., Bristow,
 M. Corbeels et al. / Europ. J. Agronomy 76 (2016) 41–53 53

K.L., Asseng, S., Chapman, S., McCown, R.L., Freebairn, D.M., Smith, C.J., 2003. Rusinamhodzi, L., Corbeels, M., van Wijk, M.T., Rufino, M.C., Nyamangara, J., Giller,
An overview of APSIM, a model designed for farming systems simulation. Eur. K.E., 2011. A meta-analysis of long-term effects of conservation agriculture on
J. Agron. 18, 267–288. maize yields under rain-fed conditions: lessons for southern Africa. Agron.
Lindquist, J.L., Arkebauer, T.J., Walters, D.T., Cassman, K.G., Dobermann, A., 2005. Sustain. Dev. 31, 657–673.
Maize radiation use efficiency under optimal growth conditions. Agron. J. 97, Saseendran, S.A., Ma, L., Malone, R., Heilman, P., Ahuja, L.R., Kanwar, R.S., Karlen,
72–78. D.L., Hoogenboom, G., 2007. Simulating management effects on crop
MacCarthy, D.S., Vlek, P.L.G., Bationo, A., Tabo, R., Fosu, M., 2010. Modeling nutrient production, tile drainage, and water quality using RZWQM-DSSAT. Geoderma
and water productivity of sorghum in smallholder farming systems in a 140, 297–309.
semi-arid region of Ghana. Field Crops Res. 118, 251–258. Scopel, E., da Silva, F.A.M., Corbeels, M., Affholder, F., Maraux, F., 2004. Modelling
Materechera, S.A., Mloza-Banda, H.R., 1997. Soil penetration resistance, root crop residue mulching effects on water use and production of maize under
growth and yield of maize as influenced by tillage system on ridges in Malawi. semi-arid and humid tropical conditions. Agronomie 24, 1–13.
Soil Tillage Res. 41, 13–24. Six, J., Conant, R.T., Paul, E.A., Paustian, K., 2002. Stabilization mechanisms of soil
McKay, M.D., Conover, W.J., Beckman, R.J., 1979. A comparison of three methods organic matter: implications for C-saturation of soils. Plant Soil 241, 155–176.
for selecting values of input variables in the analysis of output from a Sommer, R., Wall, P.C., Govaerts, B., 2007. Model-based assessment of maize
computer code. Technometrics 21, 239–245. cropping under conventional and conservation agriculture in highland Mexico.
Ngwira, A.R., Aune, J.B., Thierfelder, C., 2014. DSSAT modelling of conservation Soil Tillage Res. 94, 83–100.
agriculture maize response to climate change in Malawi. Soil Tillage Res. 143, Thierfelder, C., Mwila, M., Rusinamhodzi, L., 2013. Conservation agriculture in
85–94. eastern and southern provinces of Zambia: long-term effects on soil quality
Ogindo, H.O., Walker, S., 2005. Comparison of measured changes in seasonal soil and maize productivity. Soil Tillage Res. 126, 246–258.
water content by rainfed maize-bean intercrop and component cropping Thierfelder, C., Wall, P.C., 2009. Effects of conservation agriculture techniques on
systems in a semi-arid region of southern Africa. Phys. Chem. Earth 30, infiltration and soil water content in Zambia and Zimbabwe. Soil Tillage Res.
799–808. 105, 217–227.
Parton, W.J., Schimel, D.S., Cole, C.V., Ojima, D.S., 1987. Analysis of factors Thierfelder, C., Wall, P.C., 2010. Investigating conservation agriculture (ca) systems
controlling soil organic matter levels in Great Plains grasslands. Soil Sci. Soc. in Zambia and Zimbabwe to mitigate future effects of climate change. J. Crop
Am. J. 51, 1173–1179. Improv. 24, 113–121.
Pathak, T.B., Fraisse, C.W., Jones, J.W., Messina, C.D., Hoogenboom, G., 2007. Use of Thierfelder, C., Rusinamhodzi, L., Ngwira, A.R., Mupangwa, W., Nyagumbo, I.,
global sensitivity analysis for CropGro cotton model development. Am. Soc. Kassie, G.T., Cairns, J.E., 2014. Conservation agriculture in Southern Africa:
Agric. Biol. Eng. 50, 2295–2302. advances in knowledge. Renew. Agric. Food Syst. 30, 328–348.
Porter, C., Jones, J., Adiku, S., Gijsman, A., Gargiulo, O., Naab, J., 2010. Modeling Thioulouse, J., Lobry, J.R., 1995. Co-inertia analysis of amino-acid physico-chemical
organic carbon and carbon-mediated soil processes in DSSAT v4.5. Oper. Res. properties and protein composition with the ADE package. Comp. Appl. Biosci.:
10, 247–278. CABIOS 11, 321–329.
R Development Core Team, 2009. R: A Language and Environment for Statistical Uri, N.D., 2000. An evaluation of the economic benefits and costs of conservation
Computing. R Foundation for Statistical Computing, Vienna, Austria. tillage. Environ. Geol. 39, 238–248.
Reichert, P., Omlin, M., 1997. On the usefulness of over-parameterized ecological Waddington, S.R., Karigwindi, J., 2004. Longer-term contribution of groundnut
models. Ecol. Model. 95, 289–299. rotation and cattle manure to the sustainability of maize-legume smallholder
Ritchie, J.T., 1998. Soil water balance and plant water stress. In: Tsuji, G.Y., et al. systems in sub-humid Zimbabwe. In: Friesen, D.K., Palmer, A.F.E. (Eds.),
(Eds.), Understanding Options of Agricultural Production, Kluwer Academic Integrated Approaches to Higher Maize Productivity in the New Millennium.
Publ. and Int. Consortium for Agriciltural Systems Applications, Dordrecht, The CIMMYT, Kenya Agricultural Research Institute, Nairobi (Kenya), pp. 338–342.
Netherlands, pp. 41–53. Wallach, D., Goffinet, B., Tremblay, M., 2002. Parameter estimation in crop models:
Ritchie, J.T., Porter, C.H., Judge, J., Jones, J.W., Suleiman, A.A., 2009. Extension of an exploring the possibility of estimating linear combinations of parameters.
existing model for soil water evaporation and redistribution under high water Agronomie 22, 171–178.
content conditions. Soil Sci. Soc. Am. J. 73, 792–801.
Robert, P., Escoufier, Y., 1976. A unifying tool for linear multivariate statistical
methods: the RV-coefficient. Appl. Stat. 25, 257–265.