Introduction

Long-term experiments are crucial for determining fundamental crop, soil and ecological processes and their impacts on the environment (Johnston 1997; Körschens 2006; Leigh and Johnston 1994; Mitchell et al. 2008; Payne 2006; Poulton 1995; Rasmussen et al. 1998; Zhao et al. 2010). Long-term experiments have shown, for example, that commonly used irrigation and nitrogen fertilization practices can produce substantial N leaching out of the root zone, which in turn results in environmental degradation and substantial economic loss to the producer (Edmeades 2003; Johnston 1997). Hence, long-term experiments are particularly useful for developing beneficial management practices (BMPs) which produce economically viable crop yields while maintaining soil quality and environmental health at acceptable levels.

Classical statistics, such as analysis of variance and regression, were the most frequently used methods in analysing long term experimental field data; and Barnett (1994) illustrated how they were used to analyze the long-term field data from Rothamsted, England. Payne (2006) concluded that since regression methods (i.e., Fisher’s 5th-order Legendre polynomial, or multiple regression models between yield and climate data) could explain only about 10–40% of the variation in crop yield, many additional crop, soil and weather factors need to be included.

In recent decades, the use of long term field data to develop and evaluate dynamic simulation models make these data extremely valuable (Jenkinson et al. 1994; Körschens 2006; Leigh et al. 1994). The simulation models in turn reveal data gaps and errors, and thereby promote improved data collection and documentation (Hunt et al. 2001). In addition, dynamic crop and soil models facilitate detailed and systematic analyses by providing inexpensive, rapid and detailed estimations of crop growth and soil nutrient-water movement. By analysing model predictions, economic and environmental benefits can be compared among different crop production and soil management scenarios. After successful calibration and evaluation, a simulation model can be usefully applied to both the development of BMPs, and the analysis and forecasting of crop production systems (Jones et al. 2003; Saseendran et al. 2007).

Integrating long-term experimental data with model simulation is one of the objectives in the inter-linked group of submodels known as the DSSAT (Decision Support System for Agrotechnology Transfer) Crop System Model (Jones et al. 2003), and it has been successfully employed in long-term fertilizer, irrigation, pest management, and site-specific farming applications (Cabrera et al. 2007; Jones et al. 2003; Lobell and Ortiz-Monasterio 2006). Other models have also been used to simulate long-term experiments; e.g. the integrated Root Zone Water Quality Model (RZWQM) was used to simulate long-term management effects on crop production, tile drainage and water quality (Ahmed et al. 2007; Ma et al. 2007; Saseendran et al. 2007); the Agricultural Production Systems Simulator (APSIM) model was applied to simulate long-term cropping system with climate variability and irrigation management (Keating et al. 2003; Wang et al. 2008, 2009); the Environmental Policy Integrated Climate (EPIC) model was used to estimate crop growth, soil erosion, and nutrient loss by runoff and leaching (Izaurralde et al. 2006; Roloff et al. 1998; Williams 1995), and agroecological changes of 93-year farming transformation (Bernardos et al. 2001); and the integrated soil-crop HERMES model, was used to simulate 25 years of crop rotations on a long-term field experiment at Swift Current in south-western Saskatchewan, Canada (Kersebaum et al. 2008). A particularly useful feature of the DSSAT Crop System Model is its ability to handle site-specific crops and rotations (Bowen et al. 1998). Furthermore, incorporation of the CERES submodel into DSSAT allows simulation of soil carbon and nitrogen cycling in low-input systems, and the ability to simulate the effects of climate change on carbon sequestration and long term yield forecasting (Gijsman et al. 2002; Koo et al. 2007; Porter et al. 2009).

Maize (Zea mays) is the main field crop grown in southern Ontario, Canada, but yields vary substantially from year to year due to weather variations (Ahmed et al. 2007; Cabas et al. 2010; Drury and Tan 1995; Tan and Reynolds 2003), changes in cultivars and adoption of soil and crop management practices. Detailed characterization of the combined effects of weather, cultivars and field management on long-term maize yield and soil N dynamics are needed for the development of effective and region-specific BMPs for sustainable maize production. However, there are few simulations of long-term maize production in Canada and southern Ontario, in particular. Hence, a DSSAT-CERES-Maize simulation of a long-term continuous maize production experiment at Woodslee, Ontario was conducted. The field experiment was started in 1959 to study the long-term effects of fertilization, crop rotation and weather on crop yield (Drury and Tan 1995) and soil and water quality (Tan et al. 2002). The objectives of the study were to: (1) use the DSSAT-CERES-Maize model to simulate maize yield and soil nitrogen and water dynamics over 50 years of continuous maize production with and without fertilizer application; and (2) evaluate the DSSAT-CERES-Maize model by comparing simulated and measured maize yield and soil N data.

Materials and methods

Field experimental data

The data were collected from a 50 year long term experiment established in 1959 at the Honourable Eugene F. Whelan Experimental Farm, Agricultural and Agri-Food Canada, Woodslee, Ontario (42°13′N, 82°44′W, Elevation 186 m). The soil is Brookston clay loam (Orthic Humic Gleysol), which is a poorly drained, lacustrine soil with an average root zone texture of 28% sand, 35% silt, 37% clay, and an average pH of 6.1. Average organic carbon in the top layer (i.e., Ap horizon) was 2.0–2.5% and surface slopes are generally <0.1%. The cropping treatments included continuous maize versus a maize-oats-alfalfa-alfalfa rotation, and the fertilization treatments included no fertilization versus chemical fertilization. The fertilization treatments included both a pre-plant application of starter fertilizer as well as a sidedress application for maize at the 6 leaf stage. The starter fertilizer (16.8 kg N ha−1, 67.2 kg P2O5 ha−1 and 33.6 kg K2O ha−1) was applied as a broadcast—incorporated application to the soil 1 day before planting (Table 1). The side-dress application was a band application (2–5 cm deep) of 112 kg N ha−1 15 cm from either side of the maize when the maize was at the 6 leaf stage (Table 1). The no–fertilizer treatment did not receive any chemical or organic fertilizer over the 50 year period; and as a result, many of its soil physical and hydraulic properties (e.g. bulk density, hydraulic conductivity, organic carbon, etc.) were substantially degraded relative to the chemical fertilization treatment (Table 3). The history and treatments of the site were described previously (Bolton et al. 1970; Drury et al. 1998; Drury and Tan 1995; Tan et al. 2002). Two treatments were selected for this study, continuous corn/maize with fertilization (CC-F) and continuous corn/maize without fertilization (CC-NF). Each plot was 76.2 m long by 12.2 m wide, and tillage consisted of fall mouldboard ploughing to 0.15 m depth, then discing and harrowing in the spring prior planting. Herbicides were applied to both CC-F and CC-NF as required to control weeds.

Table 1 Field management data from 1959 to 2008 for the continuous corn with fertilization (CC-F) treatment at Woodslee, Ontario, Canada
Full size table

Planting density varied from 37,000 to 50,000 seeds ha−1 from 1959 to 1982 and 55,000 seeds ha−1 from 1983 to 1996 with a 1.0 m row spacing (Drury and Tan 1995; Gregorich et al. 2001), and it was increased to 62,000 seeds ha−1 from 1997 to 2007 with a 1.0 m row spacing. In 2008, the density was increased to 75,000 seeds ha−1 with a row spacing of 0.76 m, which is a typical corn row spacing. Maize grain yields were measured annually in the fall by harvesting 33 m lengths of 10 interior rows, and then grain sub-samples were taken to determine grain moisture. Grain yields were normalized to 15.5% grain moisture content, and maize dry yield was also calculated. Each year after harvest, maize residues were incorporated into the soil by mouldboard ploughing. Soil nitrate leaching through drainage tiles was measured during 1998–2000 (Tan et al. 2002) and soil mineral N in the 0–15 and 15–30 cm layers were measured in 2008.

Model input/output and evaluation data

Model description

The Cropping System Model (CSM) of DSSAT is a well developed process-oriented model which is capable of simulating long term rotation experiments (Tsuji et al. 1994; Jones et al. 2003). Each crop in the rotation has a separate module to simulate phenology, daily growth, development and partition biomass to leaves, stems, roots, ears and grain based on the supply and demands of soil water, carbon and nitrogen and phosphorus nutrients in response to weather and management conditions (Jones et al. 2003). The current version of DSSAT (4.5) contains 10 separate models in the DSSAT-CSM to simulate the growth of 29 different crops as well as fallow fields (Daroub et al. 2003; Hoogenboom et al. 1999; Jones et al. 2003; Jones et al. 2001; Tsuji et al. 1994). There were two soil organic nitrogen submodels in DSSAT 4.5: the CERES soil model and the Century soil model. For long-term sequence simulations, the Century-based soil model seems to predict soil organic nutrient processes and organic carbon dynamics more realistically than the CERES-based model (Gijsman et al. 2002), and hence the Century-based soil model was selected for this study. Porter et al. (2009) describe the organic carbon and carbon-mediated soil processes in the Century-based soil model. Since the crop in this study was continuous maize, the Maize submodel in DSSAT was selected.

Model input parameters

The required model inputs describe field management, daily weather, soil profile characteristics, initial soil condition, and cultivar characteristics. The primary field management inputs include crop planting date and density; fertilization rate, date and depth; tillage and irrigation systems; and the amount and method of residue incorporation. These data were assembled from publications (Drury and Tan 1995; Tan et al. 2002) and unpublished field notes, and they cover the period from 1959 to 2008. The key model inputs are summarized in Table 1.

The minimum weather input required to run DSSAT, includes daily solar radiation (MJ m−2), daily maximum and minimum temperature (°C), and daily precipitation (mm). To meet this requirement, it was necessary to use daily weather data from both the Woodslee weather station adjacent to the field site, and the Harrow weather station which is 30 km from the field site (42°02′N, 82°54′W). For 1959–1977, Harrow weather data were used, and for 1978–2008, Woodslee weather data were used, except the solar radiation for 1978–1987 which was calculated using Harrow daily sun hour data. The WeatherMan tool (Pickering et al. 1994; Wilkens 2004) in DSSAT 4.5 was used to check data, estimate missing values and calculate solar radiation from sun hours. Monthly weather data for selected years was used for cultivar calibration (Table 2).

Table 2 Monthly, growing season and annual precipitation from 1959 to 2008 at Woodslee, Ontario, Canada
Full size table

The soil profile input data included lower soil water limit or wilting point water content (cm3 cm−3), upper drainage limit or field capacity water content (cm3 cm−3), upper soil water limit or saturated water content (cm3 cm−3), dry bulk density (g cm−3), soil organic C content (wt. %), saturated hydraulic conductivity, soil pH, clay content (wt. % <0.002 mm particle size), and silt content (wt. % 0.002–0.05 mm particle size). The water content, bulk density, organic C and saturated hydraulic conductivity data were measured from intact cores (10 cm diameter by 10 cm long) collected from the 0 to 10, 10 to 20, 20 to 30 and 30 to 40 cm depths during 1999 and 2000 (unpublished data); the pH, clay and silt data were obtained from a soil survey report (Bryant et al. 1987). The data were used to construct the model soil profiles as indicated in Table 3. The DSSAT soil profile data also requires a maize root distribution/growth parameter (i.e. a factor from 0 to 1) (Stone et al. 1987), which was estimated here using the relationship: Root_growth_factor = 1.0 × exp (−0.02 LayerCentre), where LayerCentre is the depth from the soil surface to the centre of the soil layer in question (Gijsman et al. 2007). The root growth factor was set to 1.0 if the centre of the soil layer was ≤20 cm from the surface (Table 3).

Table 3 Soil profile data under continuous corn with fertilization (CC-F) and continuous corn without fertilization (CC-NF) treatments at Woodslee, Ontario Canada in 2000
Full size table

Soil profile initial conditions for the start of the simulations (May 10, 1959) were estimated by running DSSAT for 3 years (May 1 1956 to May 9 1959) using 1956–1959 weather data and the soil physical data for CC-F (Table 3), and the soil mineral N (NH4–N and NO3–N) and water content profiles measured in CC-F during year 2000. Use of the CC-F data for establishing the initial conditions of both treatments (CC-F, CC-NF) is justified on the grounds that maize was grown on all plots for 3 years 1956–1958 after tile installation (1955) to ensure that tile flow was uniform across plots before the start of the cropping and fertilizer treatments. Once the soil profile initial conditions were established, the simulations were run from May 10, 1959 to Dec 31, 2008.

Model evaluation

Model performance was evaluated by comparing simulated and measured values of maize grain yields for the entire 50 year period, soil profile N distribution for 2008, and nitrate loss in tile drainage water for 1998–2000. The comparisons were quantified via the “EasyGrapher” program (Yang and Huffman 2004).

Deviation and test statistics have been used extensively in the past to assess DSSAT performance (e.g. Rinaldi et al. 2007; Timsina and Humphreys 2006), and the merits of deviation and test statistics have been reviewed by others (e.g. Kobayashi and Salam 2000; Loague and Green 1991; Willmott 1982; Yang et al. 2000). However, each deviation/test statistic addresses only a specific aspect of a model’s performance and no single statistic provides an overall model evaluation. Hence, we employed three deviation statistics (root mean square error, RMSE, normalized-RMSE (n-RMSE), forecasting efficiency, EF) and one test statistic (coefficient of determination, R 2) in the evaluation of DSSAT-Maize. The deviation statistics were defined by:

$$ {\text{RMSE}} = \left[ {n^{ - 1} \sum\limits_{i = 1}^{n} {(s_{i} - m_{i} )^{2} } } \right]^{0.5} $$
(1)
$$ {\text{Normalized-RMSE}} = {\frac{\text{RMSE}}{{\bar{m}}}} $$
(2)
$$ {\text{EF}} = {\frac{{\sum_{i = 1}^{n} {(m_{i} - \bar{m})^{2} - \sum_{i = 1}^{n} {(s_{i} - m_{i} )^{2} } } }}{{\sum_{i = 1}^{n} {(m_{i} - \bar{m})^{2} } }}} $$
(3)

where s i and m i are the simulated and measured values, respectively, n is the number of data points, and \( \overline{\text{m}} \) is the mean of the measured data. The RMSE measures the mean discrepancy between simulated and measured values with the same unit, while the normalized-RMSE removes the unit and therefore allows comparison of model performance among variables with different units. For model efficiency, EF = 1 indicates perfect correspondence between the simulated and measured data (i.e., s = m), while EF < 0 indicates that the simulated values (s) are worse than simply using the mean of m (i.e. \( \overline{\text{m}} \)). The classical R 2 statistic (1 ≥ R 2 ≥ 0) gives the percentage of data variance accounted for by the model.

Model runs and outputs

When the DSSAT model is run, it accesses four input files in the following succession: management file, daily weather file (from 1959 to 2008), cultivar file (Table 4), and soil profile file (Table 3). The daily outputs from the model includes: crop growth (growth stage, LAI, biomass and grain yield, % N in grain and leaves), and for each soil layer, carbon, mineral N (NH4 +–N and NO3 −–N) and water content. There are also summary outputs, including: evaluation, overview and soil water balance. The outputs are readily visualized using graphical display packages within DSSAT tools (e.g. GBUILD), or by using EasyGrapher, a supporting software program (Yang et al. 2004).

Table 4 The calibrated maize cultivar coefficients using a simulated N application rate of 240 kg N ha−1 using DSSAT v4.5a
Full size table

Results and discussion

Cultivar calibration

Crop growth in the CSM- CERES-Maize model is controlled by phenologically defined growth stages, which are in turn driven by energy input in the form of growing degree-days (GDD) (Hoogenboom et al. 2003; Jones et al. 2003). Growth stages are defined in DSSAT in terms of cultivar coefficients, which are specific to both the crop cultivar and the local climate, and must therefore be individually calibrated. The CSM-CERES-Maize model uses 6 cultivar coefficients (Hoogenboom et al. 2003), three representing early growth (P1, P2 and P5), two representing grain filling (i.e., G2 and G3), and one representing the phylochron interval between successive leaf tip appearances (PHINT) (Table 4). Cultivar coefficients must be calibrated to meet the observed yield or biomass under a no stress growing condition, i.e. without water, heat or nutrient deficiencies (Boote 1999).

Four maize cultivars were calibrated over the 50 year simulation period to reflect changes in variety and planting density. As the cultivar record from 1959 to 1982 was incomplete, planting density records were used as a surrogate, i.e. 5.0 plants m−2 (1959–1965), 5.4 plant m−2 (1966–1982), 5.5 plant m−2 (1983–1996), and 6.2 plant m−2 (1997–2007). Furthermore, years 1961, 1975, 1986 and 2000 were selected for cultivar calibration because growing season precipitation from May to October of those years (i.e. 361, 467, 657 and 662 mm, respectively) indicated no significant water stress (Table 2).

Four expected average maize yield objectives were established based on maize dry yield, which was about 6,000 kg ha−1 in the 1960s with an average yield increase of 1.7% per year, 7,000–8,000 kg ha−1 in the 1970–1980 period, and about 10,000 kg ha−1 in the 2000s (Drury and Tan 1995; Tollenaar and Wu 1999). Using these statistics, expected average maize dry yields for calibration years were arbitrarily set at 6,200 kg ha−1 for 1961, 7,500 kg ha−1 for 1975, 8,200 kg ha−1 for 1986 and 10,000 kg ha−1 for 2000 with optimal fertilizer N application rates set between 120 and 200 kg N ha−1(Table 4).

Soil profile data listed in Table 3 under CC-F plot is a clay soil with water stress on crop growth and are not suitable for representing an ‘ideal’ soil in South Ontario that has less water stress for corn growth. An “ideal” soil profile for calibrating the cultivar coefficients was created by arbitrarily reducing the wilting point water content of the CC-F soil profile (Table 3) by 5, 10,15, 20 and 25% to prevent crop water stresses during the calibration years (especially during the grain filling stage). After running the CERES-Maize model for above sensitivity simulations, With the 25% reduction of the wilting point water content of the CC-F soil profile, simulated crop growth during the calibration years showed no water stress and produced good agreement between the simulated and expected average maize yields. Hence, the ideal soil profile was used for calibrating the cultivar coefficients.

A default Pioneer maize cultivar (PIO 0069) was selected for the calibration because it was well suited for the Southwestern Ontario climate and ecological zone. Four crop management files were set up for each calibration year, and a cultivar coefficient calculator program (GenCalc V2.0) developed by Hunt et al. (1993) was used to calibrate coefficients firstly by growth period (P1, P2 and P5), and secondly by filling rate (G2) and kernel numbers per plant (G3). The model’s outputs were compared with the measured data for each iteration and the coefficients were changed in sequence until the simulated and the measured values showed the best match (i.e. the smallest RMSE). After calibration of the cultivar coefficients, good agreement (i.e., percent difference) was obtained between simulated and measured maize grain yield by using the calibrated cultivar coefficients (Table 4).

Maize radiation use efficiencies, RUE, (g dry grain matter MJ−1 incident solar energy) commonly varies with cultivars, i.e., the maize RUE in Ontario, Canada might increase from 1.95 to 3.78 g MJ−1 (Tollenaar and Aguilera 1992). When we simulated 50 year CC-NF field experiment using the default RUE value (4.2 g MJ−1) in DSSAT, we found that the simulated corn yields were systematically higher than the measured data. Other studies indicated that RUE might also be affected by environmental condition, such as soil fertility (Muchow 1994; Greenwald et al. 2006), but the model might not account for this stress properly. Data in Table 3 also showed that the CC-F soil was relatively constant while that of CC-NF declined over time (i.e., bulk density increased and soil organic carbon declined), Therefore, the default RUE value (4.2 g MJ−1) in DSSAT was calibrated to a constant value of 3.9 g MJ−1 for CC-F (all years), while that for CC-NF was 3.9 g MJ−1 for 1959–1967 at which time yields declined to very low values (typically <2 t ha−1), and then re-calibrated using a RUE of 2.3 g MJ−1 for 1968–2008.

Yield evaluation

The simulated and measured maize grain yields for CC-F and CC-NF from 1959 to 2008 are displayed in Fig. 1. Both the simulated and measured yields are substantially greater for the CC-F treatment (Fig. 1a) than for CC-NF (Fig. 1b). The annual fluctuation of maize yield was reasonably simulated in CC-F for all years, and in CC-NF for 1959–1970, but not reasonably simulated in CC-NF after 1970. The similar yields between CC-F and CC-NF during 1959–1965 probably occurred because the soil condition remained roughly equivalent between the two treatments during the first 5–6 years of the experiment. After 1965, however, the lower yields in CC-NF undoubtedly reflect both the reduction in soil fertility and the decline in overall soil quality in this treatment as a result of no fertilization and reduced amounts of crop residues returned to the soil.

Fig. 1
figure 1

Comparison of simulated and measured maize grain yield from 1959 to 2008 under CC-F (a) and CC-NF (b) treatments at Woodslee, Ontario, Canada; Vertical bars were SE

Full size image

The deviations between simulated and measured annual maize grain yields were rather large for both CC-F (RMSE = 1,391 kg ha−1) and CC-NF (RMSE = 1,166 kg ha−1) (Fig. 2; Table 5), but were on the same order as those obtained by others (Sadler, et al. 2000; Ma et al. 2007). Note also that the n-RMSE values were 39% for CC-F and 82% for CC-NF, indicating a substantially larger uncertainty in simulated annual yields for CC-NF than for CC-F. Figure 2 shows, however, that the simulated annual yields are scattered more or less equally about the 1:1 line for both treatments, indicating that there were no systematic deviations in the model simulations, and that the 50-year average annual yield was simulated reasonably well for both treatments (Table 5). The imprecise simulation of annual yield fluctuations probably reflects factors and/or events that are not considered in the model, such as extreme short-duration weather events (e.g. sudden heavy rainfalls, high winds, hail damage, short-term water deficits/surpluses at critical times, etc.).

Fig. 2
figure 2

Correlation between measured and simulated maize dry yield under CC-F (a) and CC-NF (b) treatments by DSSAT model from 1959 to 2008 yield at Woodslee, Ontario, Canada

Full size image
Table 5 Statistical evaluation of annual grain yield, soil mineral N and soil nitrate loss
Full size table

The large differences between simulated and measured annual maize yields may be related to our assumptions regarding soil physical properties, cultivar characteristics and land management. DSSAT sequential simulation model assumed no changes on soil physical parameters from beginning to the end (i.e., from 1959 to 2008), but the model simulates soil nutrients without set up default parameters each year. This way, soil water content, soil carbon and nitrogen dynamics were simulated continuously from 1959 to 2008. Simulated total soil organic C decreased by 5.3 and 12.3% in CC-F and CC-NF treatments, respectively over 50 year time and simulated soil mineral N showed significant changes over time (Fig. 3), indicating that the higher soil mineral N was found ranging from 100 to 300 kg N ha−1 in CC-F plot, while soil mineral N was below 100 kg N ha−1 in the CC-NF treatment. In fact, the measured soil physical properties (especially for CC-NF), such as, soil water contents at wilting point, at field capacity and at saturated condition, bulk density and percentages of clay and silt etc. changed significantly during 50 years continuous maize cultivation, and this was evidenced from the measured data (Table 3) in 2000, while the model used constant soil physical properties over the 50 years. The calibrated “average” cultivar coefficients may not have been appropriate for every year due to weather variations (White 1998; Xue et al. 2004; Langensiepen et al. 2008). Crop growth simulations can be very sensitive to short-term changes in root zone moisture content (Lobell and Ortiz-Monasterio 2006). Maize yield is known to be effected by planting date (Anapalli et al. 2005; Soler et al. 2007), and in our simulations, a single date (May 10) had to be assumed for each year before 1983, as no planting date records were available (Table 1). Grain yield at the field site was known to be significantly related to growing season precipitation (Drury and Tan 1995). Table 2 data showed that larger variations of growing season rainfall were found to range from 233 mm in 2005–784 mm in 1989. From the fitted equation Ym = 16.982 X − 0.0126 X2 (R 2 = −0.02), where Ym = the measured corn yield and X = growing season rainfall (mm), we could obtain the maximum grain yield (Ym) of 5,722 kg ha−1 at growing season rainfall of 673 mm with a small R 2 < 0.02, indicating that crop yield was not only affected by growing season rainfall. However, the fitted equation Ys = −418 + 12.17 X (R 2 = 0.42), where Ys = the simulated corn yield, indicating that the DSSAT simulated corn yield (Ys) linearly with growing season rainfall and it did not account for other effects such as flooding at higher rainfall precisely, etc.

Fig. 3
figure 3

Simulated soil mineral N (NH4–N + NO3–N) from 1959 to 2008 under CC-F and CC-NF at Woodslee, Ontario, Canada

Full size image

Soil mineral N evaluation

The DSSAT Century-based soil model simulated vastly different soil N curves under CC-F and CC-NF (Fig. 3), which are presumed to reflect changes in soil temperature, water content, carbon content, microbiological processes, etc. between the two treatments. Due to no fertilization, the simulated soil N under CC-NF was significantly lower in both magnitude and annual variation than under CC-F. For example, the results of the simulated soil mineral N (NH4–N + NO3–N) in the 0–150 cm depth fluctuated from about 50–350 kg N ha−1 under CC-F, whereas the simulation in mineral N under CC-NF ranged from only 5–100 kg N ha−1.

Soil mineral N in the top layer

The measured and simulated soil mineral N contents in the 0–15 cm and the 15–30 cm layers during 2008 were plotted in Fig. 4a and b for the CC-F treatment, and in Fig. 4c and d for CC-NF. The simulated soil mineral N tracked the measurements reasonably well for CC-F (Fig. 4a, b), and it is noted that both simulated and measured soil mineral N reached 60 kg N ha−1 in the 0–15 cm depth after N side-dressing on June 12. Both simulated and measured soil mineral N in the CC-NF treatment were very small (<5 kg N ha−1) as a result of many years with no fertilizer N application. The n-RMSE values ranged 58–66% for CC-F and 64–89% for CC-NF, indicating large uncertainty associated with simulation of soil mineral N content, which is perhaps not surprising, given the complexity of soil N dynamics.

Fig. 4
figure 4

Measured and simulated soil mineral N content under CC-F [(a) and (b)] and CC-NF [(c) and (d)] plots in 2008 at Woodslee, Ontario, Canada; Vertical bars were SE

Full size image

Soil nitrate leaching

Nitrate leaching through sub-surface tiles not only results in environment pollution, but also represents an economic loss to the producers. Nitrate leaching was influenced by crop or fallow frequency, fertilizer N rate, precipitation, soil texture and types of crop. The simulated annual nitrate loss through tile drainage from CC-F was 2–3 times larger than from CC-NF, as might be expected (Fig. 5). For the CC-F treatment, simulated cumulative nitrate leaching through tile drainage from 1998 to 2000 showed generally good visual (Fig. 5a) and quantitative (n-RMSE = 29%) agreement with measurements, but the model consistently overestimated cumulative nitrate loss from CC-NF (Fig. 5b, n-RMSE = 160%).

Fig. 5
figure 5

Simulated and measured cumulative nitrate loss through sub-surface tiles under CC-F and CC-NF treatments from Jan 1998 to Dec 2000 at Woodslee, Ontario, Canada

Full size image

Conclusions

The CSM–CERES-Maize model was evaluated using a long-term (1959–2008) continuous maize experiment in southern Ontario on a clay loam soil. Although the model did not simulate the annual maize yield precisely (Figs. 1, 2), the model simulated crop yields were as good as other researchers that have obtained. The model simulated soil N content and leaching losses less well than yield, as might be expected, with consistent overestimates for CC-NF. The overestimation of both maize yield and soil N for CC-NF may reflect inadequate model representation of the degraded soil profile under that treatment.

We suggest that sequence analysis with DSSAT should use more than one soil profile for long-term simulations, or that modifications be made to allow the model to update soil physical profile characteristics during long-term simulations. We also recommend that short-term but large impact disturbances, such as soil erosion, high winds, flooding, hail damage and insect/weed infestations, be integrated into the model to evaluate the both the immediate and cumulative effects of short-term extreme events.