Support Vector Machines for Spatiotemporal Analysis in Geosensor Networks

Abstract

Geosensor networks are a growing source of spatiotemporal data. However, the raw data generated by these networks, as simple collections of readings from point locations, allow little analysis to be conducted directly. As such, this research presents support vector machine based methods for the extraction of estimates for the spatial extent of areal events from geosensor data and demonstrates how these results can serve as a basis for spatiotemporal analysis. Support vector machines are a recently developed class of machine learning algorithms that have seen considerable application due to their attractive generalization properties and ability to efficiently handle large datasets. While traditionally applied to classification problems, this research demonstrates how these methods can be applied to geosensor applications where decision boundaries can be interpreted as representations of the boundaries of the spatial extent of events. Once derived, these estimates are shown as capable of serving as input for existing methods for spatiotemporal analysis and enabling description of the evolution of spatiotemporal phenomena in terms of movement and deformation. As coverage of geosensor networks increases, with sensors becoming smaller and cheaper, applications of the techniques described in this research are foreseen in environmental science, public health, and security informatics.