2.3 Assessing impacts and quantifying uncertainty
2.3.2 Crop modelling
2.3.2.3 Field-scale process-based modelling
Field-scale crop models are tools aimed to simulate growth processes in plants so that technological changes and environmental effects at the farm level can be assessed (El- Sharkawy,2005;Hoogenboom et al.,1994). Initially, crop models were conceived with the objective of being perfect and comprehensive, and able to reproduce all plant functions (Affholder et al.,2012;Sinclair and Seligman,1996). However, researchers rapidly realised
Chapter 2. Literature review 33
that developing approaches that were theoretically coherent, yet different in their imple- mentation and purpose was more efficient and probably more informative. Crop modelling science has thus evolved, and this evolution has led to the development of a large number of crop models (Rivington and Koo,2011).
Field-scale models attempt to capture as many as possible key processes that occur in the plant with the greatest level of detail possible. The choice of which processes to represent in detail, and the level of detail achieved for a given process is mostly limited by the understanding of crop physiology and by available data (Boote et al.,2001;Craufurd et al.,
2013). In some cases, this choice may be driven by the purpose of the model (Affholder et al.,2012;Challinor et al.,2009c). For example, developing a crop model for hydrological and soil-related applications would require the water balance component of the model to be very well developed and calibrated (Williams et al., 1989). By contrast, developing a model to predict phenology and frost damage does not even require a water balance to be calculated (Eccel et al., 2009). Developing a model that can accurately predict agricultural yields and harvest timings at the field scale would require both processes (and many others) to be correctly parameterised (Boote et al.,1996).
Field-scale models all hold different assumptions and hence show varied predictive skill.
Jamieson et al.(1998) demonstrated that high predictive skill in yield prediction in various crop models does not imply agreement in the underlying processes (e.g. LAI evolution, biomass assimilation) both across the models and between each model and the observations (Figure 2.8). Similarly, Bachelet and Gay(1993) found that impacts of high temperatures on grain yield varied significantly (12-62 %) among four rice simulation models, but showed that more detailed models were also the most skilled ones (CERES-Rice and MACROS). The same study showed that these two crop models held significant differences in the ways they simulated the interaction between high temperature and high CO2 concentrations
(Figure 2.9) (Bachelet and Gay, 1993). In principle, however, models can be parame- terised to match observations or to match other models by varying the so-called “genetic coefficients”, a concept that has been widely practiced by crop modellers (Baenziger et al.,
2004;Boote et al., 2003).
Probably the main advantage of a field-scale model is the fact that the models can be used to make decisions that can have a direct effect on farmers. Many crop modelling studies rely on the use of a field-scale model, primarily because the prediction of the model (if
Figure 2.8: Comparison of model predictions of the time course of LAI with observations for a number of models. Hollow dots are observations. AFRCWHEAT2 ( ), CERES- Wheat (...), Sirius ( ), SUCROS2 (. . . .) and SWHEAT (− − − − −−). The horizontal lines represent a radiation interceptance of 90 %. ( ) for AFRCWHEAT2 and Sirius, (...) CERES-Wheat, SWHEAT and SUCROS2. Taken from Jamieson
et al.(1998).
the model is well parameterised and if all associated data –soils, initial conditions and weather– correctly reflect the field’s conditions) would closely reflect the response of the plants in the field (Boote et al.,1996; Easterling et al., 2003; Meinke et al., 1997). This allows the researcher making decisions at the scale that is relevant for the farmer, and allows impacting agricultural production more directly (Jones et al., 2003). Second, by modelling more processes and in more detail, the models can be used to identify a wider variety of processes that are influential to crop yields under future scenarios. Third, by being modular, most of these crop models allow some flexibility in how certain processes are modelled. For instance, some of the Decision Support System for Agrotechnology Transfer (DSSAT) models allow the use of three biomass accumulation equations (including Farquhar’s photosynthesis,Alagarswamy et al. 2006) and two evapotranspiration equations (Jones et al.,2003). Fourth, carbon balances in the plant are often modelled, thus allowing a more profound analysis of source-sinks in the plant. Fifth, the detailed incorporation of “genetic coefficients” allows the further link with actual genetic data into crop model
Chapter 2. Literature review 35
Figure 2.9: Differences in the parameterised response of two rice crop models to changes in temperature and CO2. (A) CERES-Rice, and (B) MACROS. Taken fromBachelet and
Gay(1993).
predictions (Hoogenboom and White, 2003). Although predictive skill may not improve, it bridges the gap between crop breeders, physiologists and modellers (Boote et al.,2003;
Hoogenboom and White,2003;White et al.,1996). Finally, as more data becomes available from targeted agronomic trials (Craufurd et al.,2013), it is more logical to pursue the inter- comparison and improvement of field-scale models as tools that are already well advanced for predicting plant responses to varying environment and management conditions at scales in which decisions can be made regarding changes in cropping systems.
spatial scales, a topic of high interest in the context of climate change. Despite various successful attempts to using these models at larger scales (Baron et al., 2005; Challinor and Wheeler, 2008b; Jagtap and Jones, 2002), it remains unclear as to what extent, for example, using a field-scale model at a larger scale is comparable or better than using a large-area model directly (see Sect. 2.3.2.4). Field-scale models also have a very large number of model parameters, which increases the likelihood of over-parameterisation. For example, many of the crop models in DSSAT have around 150 or more coefficients to which a given simulation is sensitive. With such large number of parameters there is an increas- ing likelihood of predicting accurate yield responses for the wrong reasons (Jamieson et al.,
1998). Another potential caveat in field-scale models is related to the concept of “genetic coefficient”, which has prevented crop modellers from quantifying parametric uncertainty, a subject widely recognised and investigated by other crop modellers (Challinor et al.,
2005d; Lobell and Burke, 2010; Tao et al., 2009), and also in climate science (Murphy et al.,2007; Stainforth et al., 2005). The belief that genetic coefficients are unique com- binations of parameters that represent a given cultivar has prevented crop modellers from quantifying parameter uncertainty in field-scale crop models. This is particularly relevant given that not all model parameters are sufficiently constrained by observed data (Beven,
2006; Challinor et al.,2009b).
For a basic operational mode, field-scale crop models require daily data for maximum, minimum temperatures, precipitation and solar radiation. However, if more complex wa- ter balance and photosynthesis equations are used, more detailed meteorological data are required (e.g. relative humidity, dew point temperature). Thus, field-scale models tend to be much more data-intensive than empirical models. Similarly, parameterising or evaluat- ing the outputs of a field-scale model requires a significant amount of field measurements, including leaf area index, biomass, crop transpiration, stomatal conductance, amongst others (Boote et al.,2013; Craufurd et al.,2013). When available measurements are lim- ited to yields and phenology, uncertainties in parameterising the models can be rather large (Adam et al.,2011;Challinor and Wheeler,2008b;Ruane et al.,2013). Field-scale models are also sensitive to a wider range of inputs to which other models are normally not sensitive (Challinor et al.,2009b;Lobell and Burke,2010). These include initial soil nutri- ent (nitrogen, phosphorous, potassium), residue and organic carbon contents, soil fertility and salinity, fertiliser input, type of crop rotations, tillage, residue incorporation during the growing season, and sowing density. In many cases, due to lack of data, researchers
Chapter 2. Literature review 37
must make assumptions of management and initial conditions, thus constraining the skill of the model’s predictions (Pathak et al.,2003;Ruane et al.,2013), and potentially being a source of uncertainty in impacts projections. Sampling of these uncertainties is possi- ble and should be a topic for future research in crop modelling, particularly under future scenarios, when initial conditions cannot be ascertained.