NONLINEAR VISUALIZATION OF GEOPHYSICAL DATASETS WITH SELF-ORGANIZING MAPS

Self-Organizing Maps (SOMs) provide a powerful, nonlinear technique to optimally summarize and visualize complex geophysical data sets using a user-selected number of “icons” or SOM states, allowing rapid identification of preferred patterns, predictability of transitions, rates of transitions, hysteresis in cycles, and other aspects of data variability. SOM analysis produces a discrete, nonlinear classification of the continuum of conditions within the input data set without prior specification (or knowledge) of the “correct” output. The utility of the SOM-based approach is demonstrated here through examples of application to a number of geophysical datasets including Antarctic sea ice extent and North Atlantic mean-sea-level pressure (MSLP).

SOM-based patterns of sea-ice extent concisely capture the spatial and temporal variability in these data (covering 1973-1996), including the annual progression of expansion and retreat, a general eastward propagation of anomalies during the winter, and subannual variability in the rate of change in extent at different times of the year (e.g., retreat in January is faster than in November). There is also often a general seasonal hysteresis, i.e., monthly anomalies during cooling follow a different spatial path than during warming.

Analysis of North Atlantic MSLP data finds a North Atlantic “monopole” roughly co-located with the mean position of the Azores High, as well as the well-known North Atlantic Oscillation (NAO) dipole between the subpolar Icelandic Low and the subtropical Azores High. Little trend is shown in December, but the Azores High increased along with the NAO in January and February over the study interval (1957-2002), with implications for storminess in northwestern Europe.

Scientific use of SOMs is increasing rapidly, particularly within the climatology community (including relating ice-core glaciochemistry from Greenland and West Antarctica to atmospheric circulation patterns), yet the power of this technique for data visualization is still to be fully exploited.