Daily HUME: Daily Homogenization, Uncertainty Measures and Extremes

全球变化不仅影响长期平均气温，还可能导致频率分布的进一步变化，特别是它们的尾部。对整个频率分布的研究是重要的，因为例如，热浪和寒潮是气象事件造成的相当大一部分发病率和死亡率的原因。每日数据集对于研究这种极端天气和气候是必不可少的，因此是具有巨大社会经济后果的政治决策的基础。要可靠地评估这种变化，需要高质量的同质观测数据。不幸的是，测量记录包含许多非气候变化，例如，由于重新定位、新的气象屏或仪器而导致的均匀性。这些变化不仅影响平均值，而且影响整个频率分布。为了提高全球每日温度记录的质量和可靠性，我们提出了一种对每日温度数据进行自动均化的方法，以校正频率分布。我们建议将同质化描述为一个优化问题，并使用遗传算法来解决它。通过这种方式，整个温度网络可以同时均质化，从而提高灵敏度，同时避免设置错误(虚假)中断。通过不直接均化每日数据，而是均化月度指数(可能是月度时刻)，可以将每月均化方法的全部能力和理解带到每日数据的均化。此外，在优化框架下，可以客观而直接地确定最优的时间改正尺度，即改正是每年(所有12个月都得到相同的改正)、每半年、每季度或每月的最佳应用。这三个方面都是新的：整个网络的同时均质化、平差自由度的客观选择和改正模式的时间平均尺度。这一新方法将被应用于国际地表温度倡议的温度数据集的均质化。如此庞大的数据集需要一种自动同质化方法。为了验证该方法，我们将生成一个具有已知非均质性的人工气候数据集。为了能够生成这样一个具有实际非均质性的验证数据集，我们需要更好地理解日常数据中非均质性的本质。因此，我们打算收集和研究平行测量(一个位置有两个设置)，这使我们能够研究一个设置被另一个替换时频率分布的变化。最后，我们将研究和量化由于同质化后残留在数据集中的持续错误而产生的不确定性，并利用这一点来提高同质化算法的精度。对于使用同质化数据的气候学家来说，不确定性的知识也是不可或缺的。

Global change not only affects the long-term mean temperature, but may also lead to further changes in the frequency distribution and especially in their tails. The study of the whole frequency distribution is important as, e.g., heat and cold waves are responsible for a considerable part of morbidity and mortality due to meteorological events.Daily datasets are essential for studying such extremes of weather and climate and therefore the basis for political decisions with enormous socio-economic consequences. Reliably assessing such changes requires homogeneous observational data of high quality. Unfortunately, however, the measurement record contains many non-climatic changes, e.g. homogeneities due to relocations, new weather screens or instruments.Such changes affect not only the means, but the whole frequency distribution.To increase the quality and reliability of global daily temperature records, we propose to develop an automatic homogenisation method for daily temperature data that corrects the frequency distribution. We propose to describe homogenisation as an optimisation problem and solve it using a genetic algorithm. In this way, entire temperature networks can be homogenised simultaneously leading to an increase in sensitivity, while avoiding setting false (spurious) breaks. By not homogenising the daily data directly, but by homogenising monthly indices (probably the monthly moments), the full power and understanding of monthly homogenization methods can be carried over to the homogenisation of daily data. Furthermore, in an optimisation framework, the optimal temporal correction scale can be determined objectively and straightforwardly, that is whether the corrections are best applied annually (all twelve months get the same correction), semi-annually, seasonally or monthly. All three aspects are new: the simultaneous homogenisation of an entire network, the objective selection of the degrees of freedom of the adjustments and of the temporal averaging scale of the correction model.This new method will be applied to homogenise the temperature datasets of the International Surface Temperature Initiative. This large dataset necessitates an automatic homogenisation method. To validate the method, we will generate an artificial climate dataset with known inhomogeneities. To be able to generate such a validation dataset with realistic inhomogeneities, we need to understand the nature of inhomogeneities in daily data much better. Therefore, we intend to collect and study parallel measurements (two set-ups at one location), which allow us to study the changes in the frequency distribution if one set-up is replaced by the other. Finally, we will study and quantify the uncertainties due to persistent errors remaining in the dataset after homogenisation and utilise this to improve the accuracy of the homogenisation algorithm. The knowledge of uncertainties is also indispensable for climatologists using the homogenised data.