Comparing modelling approaches at two levels of biological organisation - Climate change impacts on selected Natura 2000 habitats
Comparing modelling approaches at two levels of biological organisation - Climate change impacts on selected Natura 2000 habitats
复制标题
DOI：
10.1111/j.1654-1103.2011.01266.x
复制
发表时间：
2011-08-01
影响因子：
2.8
通讯作者：
Beierkuhnlein, Carl
中科院分区：
文献类型：
--
作者：
Bittner, Torsten;Jaeschke, Anja;Beierkuhnlein, Carl
问：栖息地的特征是其植物种类组成。因此，气候变化对生境的影响可以通过两种相辅相成的统计方法来评估：一种是直接通过模拟生境的气候包络，另一种是间接地通过根据生境的植物种类来模拟生境。这些方法在预测栖息地分布方面有何不同？自然保护的后果是什么？地点：欧洲。方法：利用欧洲植物地图集的数据和关于自然2000地点的报告，对欧盟生境指令定义的五种自然和半自然草原生境类型的潜在气候变化对欧洲自然保护地点分布的影响进行了模拟。我们使用随机森林(RF)和Logistic回归(GLM)对2050年的当前和潜在的未来分布进行建模。结果：在不扩散的情况下，所有栖息地预计将失去22%到93%的范围。在“无限制扩散”的情况下，几乎所有的栖息地都获得了合适的气候空间，在其当前范围的5%到100%之间。在直接栖息地方法中，两种模型算法对测试数据都有很高的判别性能，并且都得到了很好的校准。在间接物种方法中，只有GLM表现出较高的模型性能；RF模型被过度拟合。发生概率的预测在模型方法(直接和间接)之间的差异比模型算法(GLM和RF)之间的差异更大。结论：生境是复杂的实体。由于生境类型的动态性质，特别是在面对气候变化的情况下，我们建议不仅根据栖息地类型的当前定义和地图分布，而且还根据其构成要素，特别是其特有的植物物种，对生境类型的未来分布进行建模。
Question: Habitats are characterized by their plant species composition. Therefore, climate change impacts on habitats can be assessed by two complementary statistical approaches: either directly by modelling the climate envelope of the habitat, or indirectly by modelling the habitat in terms of its plant species. How do these approaches differ in their projected habitat distribution? What are the consequences for nature conservation?Location: Europe.Methods: Potential climate change impacts on the distribution of European protected Natura 2000 sites were modelled for five natural and semi-natural grassland habitat types, defined by the EU Habitats Directive, using data from the Atlas Florae Europaeae and reports on Natura 2000 sites. We used random forests (RF) and logistic regression (GLM) to model current and potential future distributions for 2050.Results: All habitats are projected to lose between 22% and 93% of their range in the 'no dispersal' scenario. In the 'unrestricted dispersal' scenario, almost all habitats gain suitable climate space, between 5% and 100% of their current range. In the direct habitat approach, both model algorithms have high discriminatory performance on test data and are well calibrated. In the indirect species approach, only GLM shows high model performance; RF models are overfitted. Projections of occurrence probabilities differ more strongly between model approaches ('direct' versus 'indirect') than between model algorithms (GLM versus RF).Conclusions: Habitats are complex entities. Because of their dynamic nature, particularly in the face of climate change, we suggest modelling the future distribution of habitat types not exclusively based on their current definitions and mapped distributions, but also based on their constituent elements, and in particular their characteristic plant species.