Steen Valerie, Sofaer Helen R, Skagen Susan K, Ray Andrea J, Noon Barry R
Department of Fish, Wildlife and Conservation Biology Colorado State University Fort Collins CO USA.
U.S. Geological Survey Fort Collins Science Center Fort Collins CO USA.
Ecol Evol. 2017 Sep 20;7(21):8841-8851. doi: 10.1002/ece3.3403. eCollection 2017 Nov.
Species distribution models (SDMs) are commonly used to assess potential climate change impacts on biodiversity, but several critical methodological decisions are often made arbitrarily. We compare variability arising from these decisions to the uncertainty in future climate change itself. We also test whether certain choices offer improved skill for extrapolating to a changed climate and whether internal cross-validation skill indicates extrapolative skill. We compared projected vulnerability for 29 wetland-dependent bird species breeding in the climatically dynamic Prairie Pothole Region, USA. For each species we built 1,080 SDMs to represent a unique combination of: future climate, class of climate covariates, collinearity level, and thresholding procedure. We examined the variation in projected vulnerability attributed to each uncertainty source. To assess extrapolation skill under a changed climate, we compared model predictions with observations from historic drought years. Uncertainty in projected vulnerability was substantial, and the largest source was that of future climate change. Large uncertainty was also attributed to climate covariate class with hydrological covariates projecting half the range loss of bioclimatic covariates or other summaries of temperature and precipitation. We found that choices based on performance in cross-validation improved skill in extrapolation. Qualitative rankings were also highly uncertain. Given uncertainty in projected vulnerability and resulting uncertainty in rankings used for conservation prioritization, a number of considerations appear critical for using bioclimatic SDMs to inform climate change mitigation strategies. Our results emphasize explicitly selecting climate summaries that most closely represent processes likely to underlie ecological response to climate change. For example, hydrological covariates projected substantially reduced vulnerability, highlighting the importance of considering whether water availability may be a more proximal driver than precipitation. However, because cross-validation results were correlated with extrapolation results, the use of cross-validation performance metrics to guide modeling choices where knowledge is limited was supported.
物种分布模型（SDMs）通常用于评估气候变化对生物多样性的潜在影响，但一些关键的方法学决策往往是随意做出的。我们将这些决策产生的变异性与未来气候变化本身的不确定性进行比较。我们还测试了某些选择是否能提高外推到变化气候的技能，以及内部交叉验证技能是否表明外推技能。我们比较了美国气候动态的草原坑洼地区29种依赖湿地繁殖的鸟类的预计脆弱性。对于每个物种，我们构建了1080个物种分布模型，以代表以下因素的独特组合：未来气候、气候协变量类别、共线性水平和阈值设定程序。我们研究了每个不确定性来源导致的预计脆弱性变化。为了评估在变化气候下的外推技能，我们将模型预测与历史干旱年份的观测数据进行了比较。预计脆弱性的不确定性很大，最大的来源是未来气候变化。很大一部分不确定性还归因于气候协变量类别，水文协变量预测的范围损失是生物气候协变量或其他温度和降水摘要的一半。我们发现，基于交叉验证表现做出的选择提高了外推技能。定性排名也高度不确定。鉴于预计脆弱性的不确定性以及用于保护优先级排序的排名结果的不确定性，在使用生物气候物种分布模型为气候变化缓解策略提供信息时，有几个因素显得至关重要。我们的结果强调要明确选择最能代表可能是生态对气候变化响应基础的过程的气候摘要。例如，水文协变量预测的脆弱性大幅降低，突出了考虑水的可利用性是否可能比降水更直接的驱动因素的重要性。然而，由于交叉验证结果与外推结果相关，因此支持在知识有限的情况下使用交叉验证性能指标来指导建模选择。
