On detection of spatiotemporal clustering

Timely detection of an event is one of the important tasks in many spatiotemporal applications. The unknown nature of the event makes the detection process challenging. Likelihood ratio (LR) methods have been widely used for identifying locations and change point times through searching the maximum evidence over temporal and spatial windows. However, the LR-based methods may not be optimal if shift parameters are unknown. This paper introduces a new test framework for spatiotemporal surveillance based on the EWMA technique. We propose a statistic that applies exponential smoothing to both temporal and spatial axes and aggregate spatiotemporal data in each scan window. The results show that the EWMA-based method could be more or less sensitive than the LR-based methods depending on the chosen smoothing parameter, size of scan window, underlying shift patterns and time of occurrence.