B1065
Title: Modeling and detecting changes in spatio-temporal processes
Authors: Gaurav Agarwal - Lancaster University (United Kingdom) [presenting]
Idris Eckley - Lancaster University (United Kingdom)
Paul Fearnhead - Lancaster University (United Kingdom)
Abstract: Changepoints have been extensively studied for time series data, but there is limited literature on detecting changes in spatial processes over time. A likelihood-based methodology is developed for the simultaneous estimation of both changepoints and model parameters of spatio-temporal processes. Contrasting to existing spatial changepoint methods, which fit a piecewise stationary model assuming independence across segments, we fit a nonstationary model without any independence assumption. To deal with the complexity of the full likelihood model, we propose a computationally efficient Markov approximation. We study the effect of such an approximation and compare our method with existing methodologies through a comprehensive set of simulation studies. The method is employed for changepoint detection and missing data prediction in daily wind speeds across different synoptic weather stations in Ireland over a period of two years.
