The tenth IMSC, Beijing, China, 2007 - International Meetings on ...

We present applications of several techniques intensively studied recently in the
machine learning community to exploratory analysis of global climate data. <strong>The</strong> considered
techniques are related to rotating principal components using various statistical criteria. <strong>The</strong>
proposed algorithms use modeling assumptions plausible for the global climate system and
they are closely related to the method called independent component analysis (ICA). <strong>The</strong>
proposed exploratory algorithms follow a common framework, which allows for finding
underlying signals with properties of interest. <strong>The</strong> algorithms are iterative, alternating between
estimating the important structure of extracted signals and using this structure to find new
signal estimates.
<strong>The</strong> proposed algorithms are applied to analysis of the daily global data from the
NCEP/NCAR reanalysis project. <strong>The</strong> first example presents how using only the assumption of
prominent interannual variability allows for finding the component capturing El Nino-Southern
Oscillation (ENSO): <strong>The</strong> time course of the most prominent component is a good ENSO index
and the corresponding spatial patterns have many features traditionally associated with ENSO.
Interestingly, the second component extracted from the dataset combining the three variables
somewhat resembles the derivative of the ENSO index. It is also possible to capture how the
effect of ENSO changes depending on the time of the year using a slightly modified technique.
Several other algorithms with applications to the same climate dataset are presented. One
technique finds a meaningful representation of the slow climate variability as combination of
trends, interannual quasi-oscillations (including ENSO), the annual cycle and slow changes in
seasonal variations. This can be achieved by a similar technique which emphasizes slow
behavior of time signals prominent on distinct timescales. Another approach is to analyze fast
weather variations and find components in which such variability has significant behavior in a
slow timescale. For example, it is possible to find fast varying components whose intensity of
variations has been slowly increasing over the last decades.
<strong>The</strong>re are lots of possible directions in which this research can be extended. For example,
one can study slow climate changes, nonlinear effects such as interactions between
phenomena (similar to ENSO-annual cycle interactions), spatially evolving phenomena,
climatic effects on ecology, economy etc. It might also be possible to learn the dynamics of
some climate phenomena for making prediction of their evolution. We developed a relevant
algorithm which allows for finding groups of components with shared dynamics and predictable
time course.
<strong>The</strong> same algorithmic framework could be used for implementing new tools for
discovering significant regularities in the climate. <strong>The</strong> new directions can be motivated by the
type of features which might emerge during exploratory data analysis or can be based on
expert knowledge.
31