Information Analysis of Spatiotemporal Data Stream–Models, Algorithms and Evaluations

Abstract

The Spatiotemporal data stream has been widely used in different applications for system surveillance, prediction, and optimization. In the past decade, the advancement of sensing and data storage technologies has made spatiotemporal data more achievable and enlarges spatiotemporal data’s scale greatly. It brings opportunities as well as challenges to spatiotemporal data steam analysis. Different spatiotemporal data stream problems have their unique methodologies, but they all share one major difficulty: high dimensionality. Some of them are data-intensive in spatial space, which needs modeling and feature extraction to reduce spatial dimensions. Others are data-intensive in temporal space, which needs temporal dimension reduction. This dissertation investigates spatiotemporal data modeling with one spatial intensive application and one temporal intensive application. The spatial intensive application is the border surveillance with Unmanned Vehicles (UVs), where multiple UVs collaboratively collect image information of a target area in real-time. Millions of pixel data are observed at each timestamp from multiple UVs. The temporal intensive application is the water distribution system (WDS), where hydraulic sensors are deployed in the underground water pipe network. Each sensor measures hundreds of thousands of hydraulic readings per day. In the UVs surveillance application, a grid-based model is proposed to aggregate UAV’s global low-resolution observation and UGVs’ local high-resolution observation, which extracts crowd dynamics information from spatially heterogeneous high dimensional data. These extracted crowd dynamics data are then processed by a proposed Bayesian dynamics model for real-time crowd tracking and prediction. These models are validated and compared with benchmarks by simulation studies and a field test. In the WDS application, a penalized free-knot B-spline model is proposed to model high dimensional temporal profile data, reducing temporal dimension from hundreds of thousands of timestamps into dozens of profile coefficients. A real-time anomaly detection model is then proposed based on these modeled profiles. This model detects system anomaly (i.e., water pipe burst) from spatiotemporal profile based on a Bayesian basis-expansion model. One simulated dummy WDS and one simulated WDS at Austin, TX, are used for model validation and comparison.