IR72 (IRRI, Philippines)

Experimental treatment and design

The popular cultivar IR72 was used in this experiment. Treatments were laid out in four replicates in a split-plot design, with N treatments as the main plots. All experiments were fully irrigated, and were kept as free from weeds, pests, and diseases as possible. Fertilizer N application rates in the calibration set ranged from 0 to 225 kg N ha–1 with different splits (Table 1). In the validation set, nitrogen application rates varied from 0 to 400 kg N ha–1 (Table 2).

Table 1. Nitrogen fertilizer application rate (kg N ha–1) and splits (kg N ha–1) in the calibration set. WS indicates the wet season and DS means the dry season, and the number following is the year. The code is used for field experiments as well as for simulation study.

Table 2. Nitrogen fertilizer application rate (kg N ha–1) and splits (kg N ha-1) in the calibration set. WS indicates the wet season and DS means the dry season, and the number following is the year. The code is used for field experiments as well as for simulation study. Transp., Tr+2w, MT, PI, Fl, and Fl+1w indicate the split application time at transplanting, 2 weeks after transplanting, mid-tillering, panicle initiation, flowering, and 1 week after flowering.

Cultural practice

Twelve-day-old seedlings grown in trays were transplanted into each subplot.

Data Collection

At the major phenological stages and in between, samples were taken from 12–14 hills. LAI and organ dry weights were measured. Sequential crop samples were taken during the growing season from 14 hills to determine LAI and biomass of green leaves, dead leaves, stems, and panicles. At harvest, yield components were measured, including individual grain weight. The phenological dates of calibration set are recorded in Table 3 for all treatments, and Table 4 presents the phenological dates in the validation set.

Table 3. Phenology development of IR72 rice under different fertilizer management practices at the IRRI farm for model calibration.

Table 4. Phenology development of IR72 rice under different fertilizer management practices at the IRRI farm for model validation.

Simulations and parameterisation

Following the general methodology in Section 2, each treatment in Tables 1 and 2 was one simulation. We simulated crop growth and development for each treatment using actual fertilizer regimes, emergence dates, seedbed durations, transplanting densities, and daily weather data. For each treatment and experiment, the same model parameters were used, except for development rates, which were treatment-specific. We know that N level may affect development rate even though this effect is not yet included in the model. The development rates associated with different treatments were derived from the actual phenology development as recorded in Tables 3 and 4. In general, however, development rates should be quite stable across environmental conditions. In fact, slightly different development rates were used for different treatments in both calibration and validation sets (Table 5), because the algorithms of current ORYZA2000 did not involve functions to express the effects of soil-water and nitrogen status on phenology development. For indigenous soil N supply, we used a value of 0.8 kg ha–1 d-1 for the dry season and 0.6 kg ha-1 d-1 for the wet season, calculated from total N uptake divided by growth duration in zero-N plots in the calibration data set.

Table 5. Phenological development rates of IR72 rice under different nitrogen fertilizer management practices. DVRJ, DVRI, DVRP, and DVRR are the development rates ((°Cd)-1) in juvenile, photoperiod-sensitive, panicle development, and reproductive phases, respectively.

Evaluation of simulated results

The standard evaluation method as introduced in Section 2.5 was used to conduct the model evaluation in this experiment. The biomass of crop organs and LAI were the components used for evaluating model performance.

Because we could not retrieve the variations in measured values of means for our data sets, we estimated standard deviations (SD) and coefficients of variance (CV) for measurements of biomass and LAI from recent experiments at IRRI. These experiments used rice variety Apo under flooded conditions with high-N and zero-N levels, in four replicates, in four wet seasons and four dry seasons from 2001 to 2003 at the IRRI lowland farm (for experimental details, see Castañeda et al 2002). Because measurements in these experiments followed the same protocols as in our data sets, we used these SD and CV values as proxies for experiments with flooded rice (Table 6).

Table 6. Standard deviation (SD, same unit as variable) and coefficient of variation (CV, %) of measured crop growth variables in experiments with flooded rice. Data calculated from six seasons of field experiments at IRRI using variety Apo under flooded conditions with zero-N and high-N inputs (120 kg ha–1 in the wet season and 150 kg ha-1 in the dry season). N is the number of data pairs.

Parameter values

The parameters of phenological development were derived from recorded phenological development of Tables 3 and 4. The development rates in Table 5 can be used in other experiments with the same rice varieties because development rates should be quite stable across environmental conditions.

Biomass and LAI (calibrated data set)

Typical examples of comparisons between simulated and measured crop growth variables are given for the wet-season experiment of 1991 in Figure 1 and for the dry-season experiment of 1992 in Table 6. In both seasons, the dynamics in biomass of leaves, stems, and panicles was simulated quite well at all levels of N ranging from 0 to 225 kg ha–1. In the wet season, simulated LAI values consistently exceeded measured values in the midst of the growing season at all N levels. In the dry season, simulated LAI exceeded measured LAI only at 0 N, whereas good fits were obtained at 180 and 225 kg N ha–1. In the other years, we also got a better fit between simulated and measured LAI at high levels of N than at low levels of N. In individual years or treatments, better results were obtained with treatment-specific lower values of specific leaf area than with the average values used in our simulations.

Figure 3 compares simulated with measured crop growth variables for all data of the calibration set.  For reference, the 1:1 line plus and minus the estimated SD of measured variables is also shown. The best results were obtained for total aboveground biomass, for which most of the data points fell between the ±SD lines of measured biomass. There is more spread in the data of leaf, stem, and panicle biomass, and more data fell outside the ±SD lines. However, the most spread is observed for LAI, for which more than half of the data points were above the ± SD lines, indicating a consistent overestimation of LAI. Figure 4 gives the simulated and measured yields and final biomass at harvest, together with the 1:1 ± SD lines. All simulated biomass values fell within or close to the 1:1 ± SD lines, whereas about 25% of the simulated yields were below the 1:1 ± SD lines.

Table 7 gives the RMSE for each treatment and experiment separately, the goodness-of-fit parameters for the dynamic crop variables of the whole data set, and the parameters for yield and final biomass at harvest of the whole data set. Some variation occurred in RMSE among treatments and years, but general patterns were consistent. The RMSE of LAI was consistently largest and that of total aboveground biomass consistently smallest. Moreover, except for LAI, the range in RMSE values for each crop variable was small. There were no relationships between RMSE values and total amount of N applied. Using the whole calibration set, Student’s t-test indicates that simulated crop growth variables were similar to measured values except for LAI (Table 7). For LAI, the slope was close to 1, but the intercept was high, which indicates a general overestimation of simulated values. The relatively low R2 reflects the large spread in the data.  The absolute RMSE and the normalized RMSE of LAI simulations were about three times greater than the typical SD and CV values of measured LAI, respectively (Table 6). The normalized RMSE of simulated aboveground biomass was similar to the CVs of measured values. However, the RMSE of simulated biomass of leaves, stems, and panicles was on average 65% higher than the SD values of the measurements. For final yield and end-of-season biomass, all goodness-of-fit parameters indicate a close fit between simulated and measured data (Table 7).

Biomass and LAI (validation data set)

Figure 5 compares simulated and measured crop growth variables against time at 0, 100, and 400 kg N ha–1. Compared with the calibration set (Figs. 1 and 2), simulated total aboveground biomass values exceeded measured values, whereas simulated biomass of leaves, stems (not shown for clarity’s sake), and panicles matched measured values well. At 400 kg N ha–1, LAI was simulated well, but, as in the calibration set, with decreasing levels of N, simulated LAI values exceeded measured values. In Fig. 6A, at 0 kg N ha–1, the transition between exponential and linear leaf area growth phases is clear. This graph suggests that the effect of N limitation during exponential growth was simulated relatively accurately, but that, in the linear phase, N limitations may have reduced specific leaf area, which the model did not simulate.

Figure 6 compares simulated and measured crop growth variables versus time with the same amount of N (300 kg N ha-1) but in different splits and application timings. When all N was applied after flowering, little was taken up and LAI and biomass remained low (Fig. 3A). Measurements were fairly well reproduced by the model, though simulated aboveground biomass and LAI exceeded measured values. When N was applied earlier in the season, LAI and biomass increased (Fig. 6B and C). In these cases, all crop growth variables were simulated well. Note again in Figure 6A and B the inflection in LAI curves, indicating the transition from exponential to linear leaf area growth.

Another graphical comparison between simulated and measured data is the three-quadrant diagram of total N uptake and yield versus N supply (Fig. 7). Three-quadrant diagrams clearly show the relationships between the amount of N supplied and the amount of N taken up by the crop (quadrant I), the amount of N taken up and yield (quadrant II), and the amount of N supplied and yield (quadrant III). In general, ORYZA2000 simulated yield fairly well (quadrants II and III), though simulated N uptake was often slightly higher than measurements (quadrant I). Simulated yields were about 2 t ha–1 higher than measured yields in the two treatments in which 150 and 300 kg N were applied from flowering onward. Because total N uptake was simulated well, the effect of this late N uptake on yield was overestimated by the model.

Figure 8 gives the simulated and measured crop growth variables for all data of the validation set, together with the 1:1 ± SD lines. Compared with the calibration set (Fig. 3), the spread in data was smaller. In the validation set, simulated total aboveground biomass exceeded measured values, which was mainly caused by the overestimation of stem biomass since the biomass of leaves and panicles was simulated quite well. Like in the calibration set, simulated LAI generally exceeded measured LAI. In Figure 4, simulated end-of-season biomass values fell mostly within or close to the ±SD lines of measured values, though most simulated values were above the 1:1 line. Likewise, simulated yields fell within the ±SD lines of measured values except for the two cases of late N application mentioned above. The RMSE is given in Table 6 for each treatment and experiment separately, while the goodness-of-fit parameters are given in Table 7 for the dynamic crop variables and for yield and final biomass at harvest. Goodness-of-fit parameters were similar to those in the calibration set for the dynamic biomass of leaves and panicles. However, Student’s t-test indicated that simulated and measured dynamic biomass of stems and LAI were not the same at the 95% confidence level.

For LAI, absolute RMSE and normalized RMSE (%) values were smaller in the validation set than in the calibration set, whereas the opposite was true for total aboveground biomass. Simulated and measured end-of-season biomass and yields do not differ statistically from each other. However, the values of , , and RMSE confirm the trend that simulated values are higher than measured values as observed in the graphical model evaluation (above). On average, normalized RMSE of crop growth variables was 84% higher than the typical CV of measured values (it was 27% higher for biomass of panicles and 140% higher for that of stems).

Reference

Castañeda AR, Bouman BAM, Peng S, Visperas RM. 2002. The potential of aerobic rice to reduce water use in water-scarce irrigated lowlands in the tropics. In: Bouman BAM, Hengsdijk H, Hardy B, Bindraban PS, Tuong TP, Ladha JK, editors. Water-wise rice production. Proceedings of the International Workshop on Water-wise Rice Production, 8-11 April, 2002, Los Baños, Philippines. Los Baños (Philippines): International Rice Research Institute. p 165-176.