DATE2018-05-16 03:21:06
IDABSTRACT20180516032106-0150
CONTACTrousi@pik-potsdam.de
PRESENTATIONORAL
INVITED0
IDSESSION1
TITLEIDENTIFYING REMOTE DRIVERS OF MEDITERRANEAN RAINFALL USING MACHINE LEARNING AND CAUSAL DISCOVERY TOOLS
AUTHORSE Rousi (1), M Kretschmer (1), J Lehmann (1), S Totz (1), D Coumou (1,2)
AFFILIATIONS
  1. Potsdam Institute for Climate Impact Research, Potsdam, Germany
  2. Potsdam Institute for Climate Impact Research, Potsdam, Germany
ABSTRACTThe Mediterranean region has been identified as one of the hotspots of climate change due to extensive drying and a warming trend, especially in summer. Winter rainfall is particularly important as it provides the main source of soil moisture for the following warm season, which is typically hot and dry. Over the last decades, winter precipitation has decreased and this trend is projected to continue in a warming climate. A better understanding of the drivers of Mediterranean climate variability is thus crucial to improve (sub) seasonal to decadal forecasting. Skillful forecasts are essential for actionable climate services and for policy-makers to mitigate climate impacts on agriculture, economy, infrastructure and societies in general. A lot of research has been done to study remote drivers of rainfall variability in the Mediterranean, such as North Atlantic sea-surface temperatures and large-scale atmospheric circulation. However, little attention has been given so far to the use of innovative techniques, such as machine learning tools and causal discovery algorithms, that are not simply based on correlation but rather try to tackle causality. Based on such methods, this work aims to better understand the causal drivers of precipitation variability in the Mediterranean. We will analyze large-scale atmospheric patterns, including the North Atlantic Oscillation (NAO), a well-known mode of variability over the Northern Hemisphere. NAO is established to be important for the Mediterranean climate variability, but usually its different spatial characteristics are not being considered. Here, specific "flavors" of the NAO, i.e. recurring spatial patterns identified by Self-Organizing Maps (Rousi et al., 2017), will be examined including their effects on Mediterranean precipitation. Moreover, we present studies on predicting winter precipitation anomalies in the Mediterranean using a cluster-based empirical forecast method, that considers the strength but also the spatial pattern of precursors. This method provides skillful empirical forecasts outperforming operational forecasting systems (Totz et al., 2017). Furthermore, causal effect networks have been used to identify Arctic-related drivers of the mid-latitude winter circulation and have given promising results (Kretschmer et al., 2017). A similar analysis is proposed here to study the causal drivers of Mediterranean precipitation.
PAGE8
STATE1