Joint Seminar: Uncovering the forced climate response using statistical learning: From understanding regional climate change to global Detection&Attribution

Internal atmospheric variability fundamentally limits short- and medium-term climate predictability and obscures evidence of anthropogenic climate change regionally and on short time scales. Statistical learning techniques can be used to better characterize climate signals against the noise of internal variability. Hence, these techniques improve our understanding of regional climate change, for example through characterizing and removing the influence of internal atmospheric variability (“dynamical adjustment”), but can be used also for detection and attribution (D&A) at the global scale.

First, we demonstrate in a climate model ensemble that statistical learning methods establish a consistent relationship between internal circulation variability and atmospheric target variables such as temperature or precipitation. An accurate estimate of the forced climate response with improved signal/noise ratio is maintained in the residuals of the prediction[1] (that is, after dynamical adjustment). We then show that observed regional climate features, such as the late 1980's abrupt winter temperature change in Europe and Switzerland can be well explained by the dynamical adjustment technique. The abrupt winter warming is explained by a few exceptionally mild winters caused by large-scale atmospheric circulation variability superimposed upon a relatively steady long-term warming trend.

Second, we show that statistical learning techniques are also well suited to detect global signals from a high-dimensional spatial pattern, for instance a global spatial field of temperature or precipitation. Here we show that the fingerprint of forced climate change can be detected in the observed global climate record at very short time scales: For example, we detect forced climate change on any individual day since around 2012, and since 2000 based on any individual year of data. The fingerprints of external forcing are extracted from large climate model ensembles, similar to traditional D&A techniques, before any observations come into play.

Overall, statistical learning techniques that characterize signals from high-dimensional climate data are useful for climate D&A at regional and global scales, and these techniques will further contribute to reducing uncertainties around internal climate variability.

[1] Sippel, S., Meinshausen, N., Merrifield, A., Lehner, F., Pendergrass, A. G., Fischer, E. M., and Knutti, R. (2019) Uncovering the forced climate response from a single ensemble member using statistical learning. Journal of Climate, 32, 5677-5699. doi:10.1175/JCLI-D-18-0882.1





13:30 Uhr


Bundesstr. 53, room 022/023
Seminar Room 022/023, Ground Floor, Bundesstrasse 53, 20146 Hamburg, Hamburg


Sebastian Sippel, ETH Zürich


Maria Rugenstein

Zur Übersicht