| † second crop sown after the harvest of the winter cereal. |
| ‡ only if severe water stress occurred. |
The highest input level represented the standard management for the crops cultivated in the area. The lowest represented an option to produce almost the same yields, but with reduced costs and labour requirements. As a consequence, the two levels were not drastically different. Nitrogen inputs did not limit yields. Water availability as determined by weather and irrigation management was the main limiting factor. Data affected by pest, disease, or weed pressure were not used for comparison with model simulations.
Daily rainfall, and maximum and minimum temperature data from meteorological stations at both locations were used as input for model simulations. Two years of radiation data were available from Modena and 3 years from Foggia. A temperature-based equation proposed by Bristow and Campbell (1984), in which the parameters were estimated according to Donatelli and Marletto (1993), was used to generate solar radiation data to complete the 6 years required for simulation. Measured soil water was used as an initial condition for model simulations.
Some additional data were used to perform a preliminary evaluation of other CropSyst features. They include seasonal water use data by crops at Foggia for 3 years, obtained from a simplified soil water balance using gravimetric sampling, and data from Modena for 4 years of soybean sown as a second crop after barley, including conventional and minimum tillage after barley harvest.
Model calibration
Model calibration requires parameter adjustment within a reasonable range of fluctuation as dictated by previous research, knowledge or experience. Following this principle, a few crop input parameters were calibrated. These parameters were adjusted within a narrow range given by the CropSyst User's manual, based on outputs of growth characteristics, patterns of water use, and minimization of differences between actual and simulated yields for a limited number of simulation trials. Other crop input parameters (crop and cultivar specific parameters) required by the model were either measured during the field experiments or indirectly determined from measurements. Weather, soil, management and initialization parameters were input as observed in the experiments
Simulations and Data Analysis
Rotations were simulated at Modena and Foggia starting 1 September 1985 and ending 31 December 1991, without reinitialization of the model during this period. Model simulation results were compared to the mean value calculated from experimental data replicates (average of 3 replicates at Modena and 2 at Foggia). Model performance was evaluated qualitatively using scatter plots of predicted vs.actual values, and statistically using indices proposed by Loague and Green (1991). These indices and their interpretation are presented in the following table.
| Root Mean Square Error |  |
|||
| Modeling Efficiency |  |
|||
| Residual Mass Coeff. |  |
= average actual yield;
n = pairs of dataRESULTS AND DISCUSSION
The field experiments used as sources were not specifically designed to provide data for model validation, hence some state variables of interest were lacking (e.g., leaf area index, biomass, etc.) throughout the growing season, which prevented testing of most sub-models to explain discrepancies between measured and simulated data. Daily solar radiation data were incomplete and required estimation. Moreover, the field data used to validate the model showed experimental error. In the following table, this error is illustrated for the crops studied as variability coefficient (C.V.).
| MODENA | |
| barley | |
| maize | |
| soybean | |
| soybean† |
| FOGGIA | |
| durum wheat | |
| sorghum | |
| sorghum† | |
| soybean† | |
| sunflower |
This is an element of uncertainty when comparing model predictions with observed values. In fact, all numerical indices calculated to evaluate model performance involve an error that adds to the error of prediction, the error of measured data in estimating the true mean corresponding to each specific treatment. The limitations in the data set suggest caution in evaluating model performance, and leave unanswered some questions on evaluation results.
Crop phenology
The agreement between predicted and actual predicted lengths in days of the periods from sowing to emergence, from sowing to flowering, and from sowing to physiological maturity is in general good. In few cases however, the error of the prediction was as much as 12 days (one case for durum wheat in Foggia). In general, the greatest errors were associated with the prediction of the phenology of the winter cereals, in spite of the ability of the model to account for vernalisation and photoperiod effects on thermal time.
Because the sowing dates were fairly constant during the 6 years of the experiment, we assume that the discrepancies between predicted and actual dates of phenological events in winter cereals was not significantly affected by photoperiod. Hence, the error is likely to be due to an inadequate representation of vernalisation.
Crop yield
Simulated yields of barley, maize and soybean at Modena are plotted vs. mean values of corresponding actual yields (Figure 1).
Figure 1. Predicted vs. actual yield data for crops grown at Modena during the years 1986-1990. Symbols correspond to years; the dashed line is the regression of predicted vs. actual data.
Figure 2 shows similar comparisons for durum wheat, sorghum, sunflower and soybean at Foggia.
Figure 2. Predicted vs. actual yield data for crops grown at Foggia during the years 1987-1991. Symbols correspond to years; the dashed line is the regression of predicted vs. actual data.
Individual years are represented by different symbols as the main objective of this validation was to test the capability of CropSyst to simulate year-to-year yield variability. Statistical analyses of these simulation results are presented in the following table.
Statistical indices to evaluate simulation results. Key: n, number of observations; O average measured yield; P average simulated yield; RMSE, root mean square error; EF, modelling efficiency; CRM, residual mass coefficient; slope, intercept and r2 of the regression predicted vs. measured yield.
| MODENA | |||||||||
| barley | |||||||||
| maize | |||||||||
| soybean/soybean(2nd) | |||||||||
| FOGGIA | |||||||||
| durum wheat | |||||||||
| sorghum/sorghum(2nd) | |||||||||
| soybean(2nd) | |||||||||
| sunflower |
At Modena, predicted barley yields were close to actual values, except in 1987, where the unusually large yields measured in the field experiment were seriously underestimated. We cannot explain this result with the limited detail available. In this analysis, data of the last three years of continuous barley, when pathogens affected yields, were not used.
The results of maize simulations show reasonable agreement with measured yields. The summer of 1989 was unusually rainy and maize was irrigated only once, compared to twice in standard management. Simulated yields fall on the 1:1 line if irrigation is suppressed from the input file, suggesting a wrong quantification of crop water requirements. However, average crop water use was similar to that expected for the crop (about 500 mm) and the water use also appeared correct for the other five years. Assuming that both crop growth and crop water use are correct, it is possible that the model attributed a higher storage efficiency of summer rains (no runoff was predicted during those simulations). Simulated soybean yields agreed more closely with actual yields), which included data for both the spring sowing and the summer sowing. In general, model performance can be considered satisfactory at Modena, with reasonable modelling efficiency values for maize and barley, and a good modelling efficiency for soybean.
At Foggia, model results for durum wheat were not satisfactory. Simulated yields showed year-to-year variability, but treatment effects (rotations in which durum wheat was included, which produced the horizontal data scatter each year) did not show up in the simulations. The main difference among rotations in environments where water is the limiting factor, such as Foggia, is the amount of residual soil water. Soil cracking during summertime, with cracks reaching 1 m depth, is a common phenomenon in these soils, which may have a significant effect on the soil water balance. Soil cracking is not accounted for in the model.
For sorghum, sown both in spring and as a second crop, simulations results were reasonable over a wide range of yields. However, there was no relation between predicted and actual yields for soybean as a second crop, in contrast with results for Modena. We have no explanation for this poor performance of the model. Model results for sunflower, show scatter around the 1:1 line, but variability in the range from 2 to about 4 t/ha was reasonably explained.
In general, simulation results for Foggia were less satisfactory than for Modena. The modelling efficiency and other indices at Foggia were reasonable for sorghum and sunflower but unacceptable for durum wheat and soybean grown as second crops. The poor results for durum wheat are of special importance , as durum wheat is the most common crop in the Capitanata plain. These results are consistent with findings at Modena for barley, indicating that simulation of winter cereals is the weakeast point of model application in the two regions. Despite the poor yield prediction of some crops at Foggia, the model predicted the soil water balance reasonably well there. The predicted values of water use by crops at Foggia for three years are plotted vs. actual values, as calculated with a simplified water balance from gravimetric sampling (see following figure).
This balance assumes no water percolation below the root zone, a reasonable assumption for this region and management. In spite of the experimental error of the methodology there was a reasonably good agreement between predicted and simulated data.
CONCLUSIONS
We have tested the capability of CropSyst to simulate different cropping systems using 6 years of data collected from rotation experiments at two locations, representative of the two largest plain areas of Italy. Simulations were performed by initializing state variables at the beginning of 6-year rotations without further reinitialization, thus constituting a severe test of the model's medium-term predictive capabilities. Data available did not allow detailed corroboration of model components and limited further analysis of situations where model performance was poor. Model predictions were evaluated taking into account that experimental data showed large variability, that solar radiation data were incomplete and required estimation, and that rotation effects other than water availability (e.g., weeds, diseases, pests, etc.) were not accounted for.
Simulations of phenological stages were in general accurate, though lacking in precision in some years with winter cereals, especially at Foggia. Accounting for vernalisation did not improve the accuracy of simulation. The phenology sub-model requires further testing with more detailed data to identify and correct possible weak points.
Model estimates of yield of maize, soybean, and barley at Modena, and of sorghum and of sunflower at Foggia, appeared reasonably accurate. CropSyst was not able to simulate soybean growth when the crop was sown as a second crop after durum wheat at Foggia. However, poor simulation of winter cereal yields proved to be the most critical limitation of the model, particularly at Foggia, and the variability in observed durum wheat yields at this location in different rotations could not be explained satisfactorily.
In general, CropSyst could reasonably well simulate a number of cropping systems, and appears to be a promising tool in agricultural systems research. A more detailed data set is needed for a more thorough model evaluation and to better identify improvements required in specific sub-components of the model for adaptation to Italian conditions.
ACKNOWLEDGMENTS
We gratefully acknowledge Dr. Keith
E. Saxton, USDA, for allowing the use of computer resources. We
also thank Mr. Roger Nelson for his assistance in preparing input
files. Cropping Systems Project, Italian Ministry of Agriculture
Paper no. 43.