Using machine learning to identify novel hydroclimate states
Using machine learning to identify novel hydroclimate states
复制标题
DOI：
10.1098/rsta.2021.0287
复制
发表时间：
2022-12-12
期刊：
影响因子：
5
通讯作者：
Cook, Benjamin I.
中科院分区：
文献类型：
--
作者：
Marvel, Kate;Cook, Benjamin I.
预计人为气候变化将改变未来的干旱风险。然而，干旱并不罕见或前所未有，如记录在树木年轮为基础的重建夏季平均帕尔默干旱严重程度指数（PDSI）。使用在这些工业化前气候重建上训练的无监督机器学习方法，我们识别出异常值：PDSI的空间模式相对于“正常”变化不寻常的年份。我们发现，在许多地区，离群值更频繁地确定在二十世纪和二十一世纪。当区域干旱地图集合并成一个单一的全球数据集时，这一趋势更加明显。根据定义，10%水平的异常值模式预计每十年发生一次，但从1950年到2000年，每十年有6年以上被确定为全球干旱地图集（GDA）中的异常值。使用观测数据集将GDA延长到2020年表明，在21世纪世纪，80%的年份都存在异常的全球干旱状况。我们的研究结果表明，在不求助于气候模型的情况下，世界更频繁地经历干旱条件，这在过去的自然气候变率的背景下是非常不寻常的。这篇文章是皇家学会科学+会议问题“人类世的干旱风险”的一部分。
Anthropogenic climate change is expected to alter drought risk in the future. However, droughts are not uncommon or unprecedented, as documented in tree-ring-based reconstructions of the summer average Palmer drought severity index (PDSI). Using an unsupervised machine-learning method trained on these reconstructions of pre-industrial climate, we identify outliers: years in which the spatial pattern of PDSI is unusual relative to 'normal' variability. We show that in many regions, outliers are more frequently identified in the twentieth and twenty-first centuries. This trend is more pronounced when the regional drought atlases are combined into a single global dataset. By definition, outlier patterns at the 10% level are expected to occur once per decade, but from 1950 to 2000 more than 6 years per decade are identified as outliers in the global drought atlas (GDA). Extending the GDA through 2020 using an observational dataset suggests that anomalous global drought conditions are present in 80% of years in the twenty-first century. Our results indicate, without recourse to climate models, that the world is more frequently experiencing drought conditions that are highly unusual in the context of past natural climate variability.This article is part of the Royal Society Science+ meeting issue 'Drought risk in the Anthropocene'.