Improving Detection of Changepoints in Short and Noisy Time Series with Local Correlations: Connecting the Events in Pixel Neighbourhoods
Code in the form of an R-package available at Github as package tjdc, see dedicated page
Abstract
Detecting changepoints in time series becomes difficult when the series are short and the observation variance is high. In the context of time series of environmental resource maps, it is often safe to assume that the abrupt events are spatially continuous, and so are the changepoints. We propose to utilise this assumption by means of hierarchical models where the changepoints are modelled using a spatial model. We demonstrate utility of the approach by constructing a Bayesian model based on the Potts model, with additional assumptions relevant to changepoint detection in national multi-source forest inventory maps. We discuss implementation issues and demonstrate the idea’s performance using a simulation study. We then apply the model to forest resource maps in order to detect felling events.
