Towards reliable extreme weather and climate event attribution

Abstract

Climate change is shaping extreme heat and rain. To what degree human activity has increased the risk of high impact events is of high public concern and still heavily debated. Recent studies attributed single extreme events to climate change by comparing climate model experiments where the influence of an external driver can be included or artificially suppressed. Many of these results however did not properly account for model errors in simulating the probabilities of extreme event occurrences. Here we show, exploiting advanced correction techniques from the weather forecasting field, that correcting properly for model probabilities alters the attributable risk of extreme events to climate change. This study illustrates the need to correct for this type of model error in order to provide trustworthy assessments of climate change impacts.

Fig. 1

Effect of ensemble calibration on an attribution case for boreal hot summers for a grid point in Sudan (12.6°N, 34.4°W). Panels a and b show the observed historical evolution of summer temperatures (black, two different datasets) and the single-model ensemble of the UK quasi-operational attribution system (HadGEM3-A) considering all forcings (red) and only natural forcings (blue) as anomalies from present day climatology (1981–2010) derived from the all forcings ensemble. Panels c and d show the probability distribution of temperature and the associated fraction of attributable risk (FAR) due to climate change (distribution red opposed to the blue one) for a 1 in 5-year event in the NAT simulation (x_EX)

Fig. 2

Reliability of event attribution experiments to simulate probabilities of high temperatures during boreal summers. The reliability measures the accuracy of simulated probabilities and a value of one denotes perfect reliability (see methods reliability assessment). Reliability is shown for a the single-model ensemble of the UK quasi-operational attribution system (HadGEM3-A), b the single-model system (HadGEM3-A) after the ensemble calibration, c the multi-model event attribution system of Climate of the 20th Century Plus project (C20C+) and d a 100-member ensemble of the weather@home project using HadCM3. The small boxes in panels a and b denote the ranked histogram (counts of the position where the observations fall in the ensemble over the historical period) for the grid point denoted in panel a over Sudan analysed in Fig. 1

Fig. 3

Effect of the model ensemble calibration on the fraction of attributable risk (FAR) to anthropogenic forcings of boreal hot summers. The change denotes the FAR calculated after the calibration minus the FAR calculated from the raw HadGEM3-A, as illustrated in Fig. 1 but on a global scale. The model probabilities are estimated by fitting a Normal distribution to all ensemble members. Stippling denotes a significant change in FAR determined by resampling at a 10% significance level