A0837
Title: A segmentation method for exploring multivariate spatiotemporal data
Authors: ShengLi Tzeng - National Chung Hsing University (Taiwan) [presenting]
Abstract: Understanding important structures within complicated multivariate spatiotemporal data presents many challenges. Common approaches treat spatial locations or time points as variables, transforming the problem into multiple time series or collections of maps. However, these transformations overlook correlations between nearby data in space and time, which tend to be more strongly correlated than more distant data pairs. Direct and cross-variogram functions can describe correlations at varying spatiotemporal distances, but assuming stationarity over space or time may be too restrictive. Building on variograms, clustering methods are integrated, including Delaunay triangulation, adjacency-based spatial continuity constraints, spectral clustering, and the k-modes algorithm. The integration segments space into connected regions and time into continuous intervals in a data-driven way, which allows for more effective capture of spatiotemporal relationships. Anomalies that may not align with overall patterns can also be easily identified.
