Weather signals with structured variance

The most prominent pattern in fast weather variations is the annual cycle. The annual oscillations of the activation (variance) structure are clearly visible, for example, in squared surface temperature anomalies in Helsinki, Finland (the plot below). Similar patterns are observed in other spatial locations as well.

The algorithm that we develop in this context tries to capture significant variance structures of fast signal variations in a given timescale of interest. For example, some climate phenomena might be characterized by slowly (e.g. in the timescale of decades) increasing or decreasing variance of fast weather fluctuations. This type of structure may not be easily observable in the data because of the dominant annual pattern but it might be there.

The derived algorithm was applied to fast surface temperature variations and some remarkable signals were obtained. When the emphasis was put on decadal timescale, some of the found components clearly exhibit slowly changing activation structure. Examples of such components are shown on this plot:

The preliminary results of this reseatch have been published in

• A. Ilin, H. Valpola and E. Oja. (2006) Extraction of Components with Structured Variance.
  In Proc. of the IEEE World Congress on Computational Intelligence, WCCI 2006, pp. 10528-10535, Vancouver, BC, Canada, July 2006.