GYGA Protocol for Model Calibration and Simulation of Yield Potential GYGA Protocol for Model Calibration and Simulation of Yield Potential

Crop simulation models: selection, calibration, use, and quality control of simulated yields

1. MODEL SELECTION

Desirable attributes of crop simulation models are summarized by van Ittersum et al., 2013 and shown in Appendix I. We argue that rather than using a single generic model globally, it is more important that a particular model has been calibrated and evaluated for the conditions to be simulated. Thus, models may differ per region or crop, as long as the models have been evaluated under those conditions and that this has been documented, preferably published in peer reviewed articles.. Preferably, the same model will be used for the same crop to simulate Yw and Yp at locations within the same geographical region such as Sub-Saharan Africa or North America.

2. APPROACH FOR MODEL CALIBRATION

Different crop cultivars are planted across locations, hence, it is necessary to calibrate crop models to account for differences in crop phenology and growth-related factors (Grassini et al., 2015).. Simulations will aim to portray most recently released high-yield crop cultivars (i.e., best locally adapted cultivars), grown in pure stands. Elaborated model calibration is preferred, but if not possible simple calibration can be followed. Data requirements for both calibration types are:

  1. Elaborated calibration of phenology and growth-related model parameters (for more details see Appendix II): requires field experiments where crops were grown without evident nutrient limitations and no incidence of biotic adversities and where all weather, soil, and management data required to run the field-year specific simulation are available (Appendix III). If such experiments are not available for a specific country or region within a country, we may use crop growth data from experiments in which crops are grown with optimal management for very similar regions in terms of climate and soils—hopefully from the same climate zone (Van Wart et al., 2013a).
  2. Simple (only phenology) calibration: if (1) is not possible, use the methodology in van Wart et al., 2013b to calibrate the simulated crop phenology for a given reference weather station (RWS) buffer. Briefly, model coefficients (related to phenology) are optimized until the simulated physiological maturity matches the actual physiological maturity date (or harvest date – see Appendix IV) reported by the country agronomist. This phenology calibration is preferably done based on the buffer-specific weather data and on sowing, anthesis (if available) and maturity dates specified by the country agronomist. In those countries or regions where calculated sowing-to-maturity growing-degree days (GDD) vary little among crop-buffer zones, the same GDD value can be used for all crop-buffer zones. This procedure should also be followed for calibration of the anthesis dates. If actual anthesis dates are not reported by country agronomists, these can be determined using generic rules (for example, in maize the GDD before and after silking are typically the same) or by data reported in the literature. In absence of experimental data, crop growth-related model parameters can be retrieved from the literature and/or previous modelling studies (e.g., van Heemst, 1988; Bouman et al., 2001).

3. SIMULATION OF YIELD POTENTIAL FOR A RWS BUFFER

Required inputs: For each RWS buffer, the following model-specific information is required to run long-term simulations of yield potential (Yp) or water-limited yields (Yw): long-term (>10 years) daily weather data, soil properties for each dominant soil type, cultivar-specific model parameters (see Section 2), and crop management data (sowing date or sowing rule and plant population density)

Number of simulations: For each RWS buffer x crop x water regime combination, the number of simulations of Yp (or Yw) will be equal to the number of soil type x crop cycle combinations, multiplied by the number of available years of weather data. 

Initialization of model runs: the sowing window for each RWS buffer is determined based on expert knowledge from country agronomists. Simulations are based on a fixed calendar date within that window or, alternatively, a flexible date that depends on local sowing rules (e.g., the first day within the sowing window when a 20-mm precipitation event within 7 days occurs). The approach to follow is determined for each cropping system.  For example, maize sowing in the United States occurs within a relatively fixed, narrow time window, whereas sowing starts at the onset of the rainy season in the case of Sub-Saharan Africa. A key issue for the model initialization is how to estimate soil water status at the time of sowing. Ideally, crop simulation models simulate the entire crop rotation and the soil water balance during the non-growing season. To do so the entire rotation must be simulated, considering the different crops over the different years. However, this approach cannot be followed when models cannot simulate the crop rotation, or specific crops within the crop rotation, or fallow period, or when different crop simulation models are used for different crops. For these cases, an alternative method will be used to estimate the soil water content at sowing time: for each simulated crop-season, a soil water balance (embedded or not in the crop simulation model) will be initialized some time (2-3 months) before planting date, assuming a reasonable fixed initial soil water content (50% of available soil water or determined by expert opinion). A constant soil water content at planting time for all simulated years will in general not be applied, except for cropping systems where rainfall during fallow period is sufficient to replenish the entire root zone (e.g. wheat in The Netherlands) or in the case of lowland rainfed rice.

Simulated grain yield: For each harvest year, the corresponding Yp and Yw will be reported at standard commercial moisture content (see Protocol for Actual Yield and Gap Determination).

4. QUALITY CONTROL (QC) OF SIMULATED YIELDS

The following QC will be followed to check consistency in the simulated yields:

  1. Check for average Yp or Yw average actual yield;
  2. Check for average yield gaps < 20% of Yw or Yp;
  3. Check for cases with Yw similar to Yp and/or low CVs (<5%) in water-limited environments;
  4. Check for Yw or Yp that are far beyond biophysical limits (maize: 22 t/ha, wheat, sorghum & millet: 15 t/ha, soybean: 8 t/ha). Boundary functions relating crop yield to water availability can also be used as aid (see Appendix III);
  5. Check for low values of Yp (e.g. < 3 Mg ha-1) and Yw (e.g. < 1 Mg ha-1) and respective CVs >30% and >100%.
  6. Check for simulated yields for particular locations/years that look ‘suspiciously' lower or higher than for the rest of the sites/years

If any of the above conditions is detected:

  1. Verify that the crop is actually being grown in the RWS buffer though consultation with country agronomists. If not, the RWS buffer should be eliminated for the particular crop.
  2. Re-check underpinning weather, soil, management, model parameters, and actual yields.
  3. If there is a reason to believe that there may be a misspecification in the reported sowing date, cultivar maturity, soil depth, etc., a targeted sensitivity analysis may be required to quantify the degree of uncertainty in the estimation of Yw or Yp. In rainfed systems, special attention should be put in checking that the crop season is congruent with the rainy season.
  4. Independent verification of ‘suspicious' weather data in a given location can be performed by re-running the simulations for a given RWS buffer based on weather data from a contiguous weather station located within the same climate zone.
  5. The assumption of homogenous weather, management, soil, and actual yield may not hold for RWS buffer that are too fragmented. Hence, fragmented RWS buffer can be eliminated if Yp, Yw, Yg or/and Ya look suspicious. The same may apply to fragmented (non-contiguous) CZs simulated with one RWS. Here, the parts of the CZ for which no RWS have been simulated can be eliminated if Yp, Yw, Yg or/and Ya look suspicious.

 

APPENDIX I – DESIRABLE ATTRIBUTES OF CROP SIMULATION MODELS

Desired attribute

Explanation

Daily step simulation

Simulation of daily crop growth and development based on weather, soil, and crop physiological attributes

Flexibility to simulate management practices

Key management practices include: sowing date, plant density, cultivar maturity, and irrigation

Simulation of fundamental physiological processes

Simulation of key physiological processes such as crop development, net carbon assimilation, biomass partitioning, crop water relations, and grain growth

Crop specificity

Should reflect crop-specific physiological attributes for respiration and photosynthesis, critical stages and growth periods that define vegetative and grain filling periods, and canopy architecture

Minimum requirement of crop ‘genetic' coefficients

Minimum requirement of crop-site ‘genetic' coefficients, such as maximum leaf area index, date of flowering, etc.

Validation against data from field crops that approach YP and YW

Comparison of model outcomes (grain yield, aboveground dry matter, crop evapotranspiration) against actual measured data from field crops that received management practices conducive to achieve YP (irrigated) or YW (rainfed crops)

User friendly

Models embedded in user-friendly interfaces, where required data inputs and outputs can be easily visualized, and with flexibility to modify default values for internal parameters

Full documentation of model parameterization and availability

Publicly available models, published in the peer-review literature, with full documentation, and with reference to data sources for  internal parameter values

Source: van Ittersum et al., 2013
 

 

APPENDIX II – GUIDANCE FOR ELABORATED MODEL CALIBRATION

The steps in the model calibration are the following (with some variation between different models):

  1. Calibrate phenology-related parameters by changing the model parameters that determine growing degree days (GDD), sensitivity to photoperiod, and vernalization requirements so that simulated dates of flowering and maturity are within ±15% of the observations. Weather data for such calibration should be retrieved following the GYGA Protocol for Weather Data.
  2. Starting with default growth-related model parameters, verify yield simulations and the cumulative light interception, leaf area index course and total biomass production, using the following quality control rules:
  1. If simulations are within ±15% of the experimental calibration yield data values, then stop further calibration
  2. If conditions in (a) are not met, other model parameters related with leaf area, biomass partitioning and/or potential grain size can be amended within plausible ranges (±20%).
  3. If after (b) the yield level still deviates more than ±15%, the experimental data should be carefully inspected to detect the reasons for deviations between simulated and measured yields and phenology (e.g., small plots, poor agronomic management, inappropriate sampling for yield determination, infrequent scouting of the crop to determine phenology, etc. - see Appendix III). If there is strong evidence to suspect lack of accuracy of yields and/or crop, the experimental data should be excluded from the calibration. If experimental data seem reliable, respiration and photosynthesis parameters can be changed within plausible ranges that are no more than ±10% of published ranges for these parameters, but only as a last resort, and if empirical evidence to support these changes exist.

 

APPENDIX III - GUIDELINES TO SELECT FIELD EXPERIMENTS FOR MODEL PARAMETERIZATION

Phenology parameters: whenever available from field experiments, recorded phenological stages at experimental sites can be used to parameterize the model coefficients related to phenology as long as the weather data from a local meteorological station (situated within the same climate zone) and phenological data are available (including dates of sowing, emergence, flowering and physiological maturity). More generic sowing calendars can supplement, but not replace, site-specific phenology data for model calibration.

Growth- and yield-related parameters: experiments that can be used for model calibration should have been carried out under optimal growing conditions. Experiments that received sub-optimal management practices should be avoided when calibrating growth- and yield-related model parameters. All data required to run the field-year specific simulations need to be available including: (i) local weather, soil parameters (site-specific data or derived from soil databases), management practices followed by the researchers (sowing date, plant population density, cultivar maturity), and initial soil water content (either measured or estimated through a soil water balance). Desirable measured variables for calibration of growth- and yield-related parameters include: aboveground biomass, grain yield, harvest index, and leaf area index. Things to check when screening suitable field experiments data are: yield level, discard field experiments in which yields are close to average farmer's yields in low-input, low-yield cropping systems), water supply, nutrient inputs amount (discard those where nutrient input levels were clearly not adequate for maximum yields), incidence of biotic factors (ideally, the crops should have received prophylactic applications to guarantee no yield reduction due to pests, diseases or weeds), field plots size (don't include yield data from small plots and without replications; preferably do not consider plots smaller than 10 m2 and/or without replicates), and varieties (experiment should use best locally adapted hybrids or closest best option). Boundary functions relating crop yield to water availability, complemented with expert opinion, can also be used as an aid for screening the quality of the experimental data (see in Fig 4 from van Ittersum et al., 2013).

 

APPENDIX IV - HARVEST VERSUS PHYSIOLOGICAL MATURITY DATES

In large-scale, mechanized commercial farming, harvest takes place when grain moisture content reaches a certain level at which mechanical harvest is possible and drying costs are minimized. Therefore, harvest can take place up to 4 weeks after the crop has reached physiological maturity. In these cases, using harvest date as a proxy for physiological maturity can lead to a large bias in the simulated yield. Hence, physiological maturity needs to be retrieved from cultivar total GDD or relative maturity and not from harvest date. If this information is not available, experts opinion or published data can be used to estimate the GDD from physiological maturity to harvest date. Subsequently, these GDD can be used to derive the physiological maturity date based on the reported harvest date.

 

REFERENCES

Bouman, B.A.M., Kropff, M.J., Tuong, T.P., Wopereis, M.C.S., ten Berge, H.F.M., van Laar, H.H. 2001. ORYZA2000: Modeling Lowland Rice, International Rice Research Institute, 2001.

Grassini, P., Van Bussel, L.G.J., Van Wart, J., Wolf, J., Claessens, L., Yang, H., Boogaard, H., de Groot, H., Van Ittersum, M.K., Cassman, K.G. 2015. How good is good enough? Data requirements for reliable crop yield simulations and yield-gap analysis. Field Crops Research 177, 49-63.

van Heemst, H.D.J. 1988. Plant data values required for simple crop growth simulation models: review and bibliography. Simulation Report CABO-TT Nr. 17. CABO and Department of Theoretical Production Ecology, Agricultural University, Wageningen University. Simulation report CABO-TT no. 17.

Van Ittersum, M., Cassman K.G., Grassini, P., Wolf, J. Tittonell, P., Hochman, Z.  2013.  Yield gap analysis with local to global relevance—A Review. Field Crops Research. 143, 4-17.

Van Wart J, van Bussel, L.G.J., Wolf, J., Licker, R., Grassini, P., Nelson, A., Boogaard, H.,  Gerber, J., Mueller, N.D., Claessens, L., van Ittersum, M.K.,  Cassman, K.G. 2013a. Use of agro-climatic zones to upscale simulated crop yield potential. Field Crops Research. 143, 44-55.

Van Wart, J., Kersebaum, C.K., Peng, S., Milner, M., Cassman, K.G. 2013b. Estimating crop yield potential at regional to national scales. Field Crops Research. 143, 34-43.