Concerns over climate change are motivated in large part because of their impact on human society. Assessing the effect of that uncertainty on specific potential impacts is demanding, since it requires a systematic survey over both climate and impacts models. We provide a comprehensive evaluation of uncertainty in projected crop yields for maize, spring and winter wheat, rice, and soybean, using a suite of nine crop models and up to 45 CMIP5 and 34 CMIP6 climate projections for three different forcing scenarios. To make this task computationally tractable, we use a new set of statistical crop model emulators. We find that climate and crop models contribute about equally to overall uncertainty. While the ranges of yield uncertainties under CMIP5 and CMIP6 projections are similar, median impact in aggregate total caloric production is typically more negative for the CMIP6 projections (+1% to −19%) than for CMIP5 (+5% to −13%). In the first half of the 21st century and for individual crops is the spread across crop models typically wider than that across climate models, but we find distinct differences between crops: globally, wheat and maize uncertainties are dominated by the crop models, but soybean and rice are more sensitive to the climate projections. Climate models with very similar global mean warming can lead to very different aggregate impacts so that climate model uncertainties remain a significant contributor to agricultural impacts uncertainty. These results show the utility of large-ensemble methods that allow comprehensively evaluating factors affecting crop yields or other impacts under climate change. The crop model ensemble used here is unbalanced and pulls the assumption that all projections are equally plausible into question. Better methods for consistent model testing, also at the level of individual processes, will have to be developed and applied by the crop modeling community.

人们对气候变化的担忧在很大程度上是由于其对人类社会的影响。要求评估不确定性对特定潜在影响的影响，因为这需要对气候和影响模型进行系统的调查。我们使用一套九种作物模型以及针对三种不同强迫情景的多达45种CMIP5和34种CMIP6气候预测，对玉米，春小麦，冬小麦，水稻和大豆的预计单产提供了不确定性的综合评估。为了使此任务在计算上易于处理，我们使用了一组新的统计作物模型仿真器。我们发现，气候和作物模型对总体不确定性的贡献大致相同。尽管CMIP5和CMIP6预测下的产量不确定性范围相似，与CMIP5预测（+ 5％至-13％）相比，CMIP6预测（+ 1％至-19％）对总热量总产量的影响中位数通常更不利。在21世纪上半叶，对于单个作物而言，整个作物模型之间的传播通常比气候模型中的传播更广泛，但我们发现作物之间存在明显差异：在全球范围内，小麦和玉米的不确定性主要由作物模型主导，而大豆和玉米不确定性则由作物模型主导。水稻对气候预测更为敏感。全球平均变暖非常相似的气候模型会导致完全不同的总体影响，因此气候模型的不确定性仍然是造成农业影响不确定性的重要因素。这些结果表明，大集合方法的实用性可以全面评估影响作物产量或气候变化下其他影响的因素。这里使用的农作物模型集合是不平衡的，并提出了所有预测都同样合理的假设。作物建模社区必须开发和应用用于一致性模型测试的更好方法，也可以在单个过程的级别上。