On the use of self‐organizing maps for studying climate extremes

Abstract

Understanding how climate extremes are sensitive to a changing climate requires characterization of the physical mechanisms behind such events. For this purpose, the application of self‐organizing maps (SOMs) has become popular in the climate science literature. One potential drawback, though not unique to SOMs, is that the background synoptic conditions represented by SOMs may be too generalized to adequately describe the atypical conditions that can co‐occur during the extreme event being considered. In this paper, using the Australian region as a case study, we illustrate how the commonly used SOM training procedure can be readily modified to produce both more accurate patterns and patterns that would otherwise occur too rarely to be represented in the SOM. Even with these improvements, we illustrate that without careful treatment, the synoptic conditions that co‐occur during some types of extreme events (i.e., heavy rainfall and midlatitudinal cyclone occurrence days) risk being poorly represented by the SOM patterns. In contrast, we find that during Australian heat wave events the circulation is indeed well represented by the SOM patterns and that this application can provide additional insight to composite analysis. While these results should not necessarily discourage researchers seeking to apply SOMs to study climate extremes, they highlight the importance of first critically evaluating the features represented by the SOM. This study has provided a methodological framework for such an evaluation which is directly applicable to other weather typing procedures, regions, and types of extreme events.

The Influence of Internal Climate Variability on Projections of Synoptically Driven Beijing Haze

“…To characterize regional circulation, we cluster 500 mb GPH anomalies into typical circulation patterns (“nodes”) using self‐organizing maps (SOM, Kohonen, 2001). SOM clustering has been widely used to analyze atmospheric circulation (Cassano et al, 2007; Gibson et al, 2017; Horton et al, 2015); further details are provided in Text S1 and Figure S1. Previous analyses examining circulation on haze days in Beijing (Chen & Wang, 2015; Li et al, 2018) have not objectively analyzed typical circulation patterns over all days, and SOM clustering provides an independent metric to assess multiple circulation patterns and how they affect air quality.…”

The Continuum of Drought in Southwestern North America

“…SOMs differ from traditional clustering algorithms because the maps output from the analysis are organized in a grid structure or topology, the size and shape of which is user defined. This grid structure can be of any dimensionality but is typically twodimensional for SOM applications to patterns of climate variability (e.g., Hewitson and Crane 2002;Johnson et al 2008;Gibson et al 2017).…”

“…We have systematically tested smaller (e.g., 2 3 2) through larger (e.g., 8 3 8) SOM structures. Different SOM grid shapes (e.g., square or rectangular) and sizes (e.g., 4 or 5 rows or columns) produce relatively consistent results because they group different phases of the relevant climate modes in different corners (e.g., Johnson et al 2008;Gibson et al 2017).…”

Variability of soil moisture proxies and hot days across the climate regimes of Australia

“…Heat waves have been linked to several modes of climate variability, which influence large-scale dynamical and circulation processes and subsequently changes in moisture availability [Cai et al, 2009;Pezza et al, 2011;Gibson et al, 2017a]. Modes of variability such as the El Niño-Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD), and the Southern Annular Mode (SAM) have been shown to strongly affect the rainfall in Australia [Jones and Trewin, 2000;Cai et al, 2009;Williams and Stone, 2009] and subsequently soil moisture (SM), which has been identified to lead to extreme heat wave events [Jones and Trewin, 2000;Cai et al, 2009;Perkins, 2015;Perkins-Kirkpatrick et al, 2016].…”