2019 REUs presented at the CERF Conference in Mobile, AL
On information criteria for dynamic spatio-temporal clustering
In A. Banerjee et al. (Eds.) Proceedings of the 6th International Workshop on Climate Informatics: CI 2016. NCAR Technical Note NCAR/TN-529+PROC, p. 5-8. DOI: 10.5065/D6K072N6
Modern climate data sets, including paleo-reconstructions, long-term weather monitoring records, and remote sensing data, contain a wealth of space-time information that leads to a variety of challenges related to data storage, management, and analysis. This has sparked an interest in dynamic space-time clustering algorithms that are particularly suitable for the analysis of large data streams. The trend-based clustering algorithm TRUST allows segmentation of space-time processes in real time, but requires the user to set multiple tuning parameters, and this step is usually performed in a subjective manner. Here we propose a data-driven automatic approach to simultaneously select the tuning parameters based on a penalized loss function. We focus on the two most important parameters of the TRUST algorithm, which define short-term closeness of observations across locations and long-term persistence of such closeness within an analyzed time window. We demonstrate the performance of the enhanced clustering procedure using simulated time series, and illustrate its applicability using long-term records of water temperature in Chesapeake Bay.