| 英文摘要: | Increasing use of regionally and globally oriented impacts studies, coordinated across international modelling groups, promises to bring about a new era in climate impacts research. Coordinated cycles of model improvement and projection are needed to make the most of this potential. Climate impacts ensembles, usually comprising multiple impact models, are a promising tool for projecting future crop productivity1 and increasing coordination between international modelling groups, evident in model intercomparison programmes (MIPs), is producing high-profile multi-model studies2. An increasing number of these studies are global in extent, whereas model accuracy and data quality are often better at local to regional scales. Here, we explore the implications of this trend for the design and coordination of future studies. We develop recommendations based on the assertion that a single-model intercomparison study, if it is to avoid being unwieldy, can focus on either projecting impacts, or on model evaluation and model intercomparison, but not both. Further, we assess the suitability of global versus regional studies for achieving each of these aims. Although our analysis is presented for agriculture, it applies to a range of climate impacts. We define global studies as those with full global coverage, and regional studies as those with limited geographic extent, such as a country or province. We also include in the latter the modelling of specific fields (that is, local studies), because the ultimate aim of local-scale studies is often to draw conclusions for the region. The models used for the different studies may be the same, although in practice they often differ in complexity. The value of multi-model impacts assessments in quantifying uncertainty is increasingly well documented3. However, we cannot simply take our cue from the larger body of work on climate ensembles, because impacts ensembles are different: they involve calibration towards a small subset of variables that may depend on the output variable of interest (for example, crop yield), as opposed to seeking to reproduce a broad set of properties of a closed system. For example, crop models are usually used primarily to simulate yield, which is only one of the many aspects of crop growth and development. Climate models, in contrast, are assessed on their representation of rainfall, temperature, wind (jet streams, monsoon circulations), ocean properties and a host of other physical properties. Assessment of multiple properties of impact models is less advanced. This is not least due to differences in model structure constraining the identification of comparable properties, and difficulties in obtaining adequate data, particularly at regional scales. Although these problems are not insurmountable, the relative lack of progress means that crop models are prone to often unknown compensation of errors, making on-going assessment of causal relationships in our impacts models particularly important. The same issue arises in other impacts models, for example modelling hydrology4 and tree distribution5. Unpicking this compensation of errors is intractable in practice, because it involves separating calibration from tuning, and we do not often have the data to do this adequately. Thus a model can never be truly 'validated' for future use, only continually evaluated in the light of the most recent data. Thankfully, model evaluation is becoming increasingly coordinated amongst model groups, and increasingly sophisticated. Progress has been facilitated by greater international coordination, for example, through the Agricultural Model Intercomparison and Improvement Project (AgMIP)6. The issue of compensation of errors is illustrated by a recent inter-comparison of 27 wheat simulation models, where parameter calibration led to a greater improvement in yield error than for any other variables, including leaf-area index, harvest index and cumulative evapotranspiration2. Figure 1 presents further evaluation of the calibration procedure conducted for that study. There is no clear relationship between the total number of genotypic parameters — which can be taken as a proxy of model complexity — and the relative error of either harvest index or grain yield (Fig. 1a,c). This result suggests that the models have more degrees of freedom than can be constrained by experimental data. Subsequent calibration of the models using experimental data led to significant reduction of model error, although this improvement (y axis Fig. 1b,d) was generally greater for yield than for harvest index, suggesting some compensation of errors. However, there was no relationship between the number of calibrated parameters and the reduction of model error (Fig. 1b,d); that is, no evidence of model over-tuning. Detailed comparisons of a range of model variables are needed if we are to determine the nature of the compensation of errors — that is, the extent to which the models are getting the right answer for what is, in part at least, the wrong reason. Multi-variable impacts studies could facilitate assessments of crop sustainability (through, for example, nitrogen and water use) and crop quality (through, for example, grain protein or mycotoxin concentration).
Affiliations
-
University of Leeds, School of Earth and Environment, Institute for Climate and Atmospheric Science, Leeds LS2 9JT, UK
-
International Centre for Tropical Agriculture (CIAT), A.A. 6713, Cali, Colombia
-
INRA UMR1095 GDEC, 5 Chemin de Beaulieu, Clermont-Ferrand, F-63039 France
-
University of Florida, Institute of Food and Agricultural Sciences, Agricultural & Biological Engineering Department, 221 Frazier Rogers Hall, Gainesville, Florida 32611-0570, US
-
International Livestock Research Institution, CCAFS 16 Mentone Terrace, Edinburgh EH9 2DF, UK
-
University of Bonn, Institute of Crop Science and Resource Conservation, Katzenburgweg 5, Bonn, D-53115 Germany
|
