André Skupin* - San Diego State University
Ninghua Wang - San Diego State University
Trent W Biggs - San Diego State University
The tri-space of geographic location, attributes, and time has been at the heart of spatio-temporal data mining efforts for many years, though frequently in an implicit form. It has recently been suggested that this tri-space could be more explicitly conceptualized in a multi-dimensional framework, to which all loci, attributes, and temporal elements of a given data set contribute. That framework can then guide the systematic computational and visual reexpression of tri-space data as a series of high-dimensional spaces, aiding researchers in the pursuit of hidden patterns and relationships.
Elements of this approach are applied to a 20-year data set of satellite observations of snow water equivalent (SWE) for the Northern Hemisphere. With more than 90,000 grid cells and observations aggregated at 8-day intervals, this sizable data set of almost 80 million SWE values is a well-suited test case for the proposed tri-space framework. With the self-organizing map (SOM) method at the heart of a dimensionality reduction and classification approach, we aim to identify decadal snow regimes and inter-annual transitions. After first confirming distinct overall SWE patterns - including those associated with mountainous, coastal, and continental regions – the study proceeds to perform transition analysis for a subset of geographic cells, tracing their movement through a 46-dimensional time space over the course of 19 years. The presentation will also reflect on the influence of various pre-processing options on tri-space modeling, including normalization of attributes according to geographic or temporal segmentation.