Analyzing Global Climate System Using Graph Based Anomaly Detection

Thursday, 18 December 2014: 9:30 AM
Kamalika Das1, Saurabh Agrawal2, Gowtham Atluri2, Stefan Liess2, Michael Steinbach2 and Vipin Kumar2, (1)University of California Santa Cruz, UARC, Santa Cruz, CA, United States, (2)University of Minnesota Twin Cities, Minneapolis, MN, United States
Climate networks have been studied for understanding complex relationships between different spatial locations such as community structures and teleconnections. Analysis of time-evolving climate networks reveals changes that occur in those relationships over time and can provide insights for discovering new and complex climate phenomena. We have recently developed a novel data mining technique to discover anomalous relationships from dynamic climate networks. The algorithms efficiently identifies anomalous changes in relationships that cause significant structural changes in the climate network from one time instance to the next. Using this technique we investigated the presence of anomalies in precipitation networks that were constructed based on monthly averages of precipitation recorded at .5 degree resolution during the time period 1982 to 2002. The precipitation network consisted of 10-nearest neighbor graphs for every month's data. Preliminary results on this data set indicate that we were able to discover several anomalies that have been verified to be related to or as the outcome of well known climate phenomena. For instance, one such set of anomalies corresponds to transition from January 1994 (normal conditions) to January 1995 (El-Nino conditions) and include events like worst droughts of the 20th century in Australian Plains, very high rainfall in southeast Asian islands, and drought-like conditions in Peru, Chile, and eastern equatorial Africa during that time period. We plan to further apply our technique to networks constructed out of different climate variables such as sea-level pressure, surface air temperature, wind velocity, 500 geo-potential height etc. at different resolutions. Using this method we hope to develop deeper insights regarding the interactions of multiple climate variables globally over time, which might lead to discovery of previously unknown climate phenomena involving heterogeneous data sources.