|
Abstract:
|
For large, high-dimensional spatio-functional data, we develop a general spatial partitioning model under Bayesian framework for (1) identifying spatially contiguous clusters (subregions) that assume different underlying spatial processes to address the nonstationarity issue; (2) improving the computational feasibility via parallel computing over subregions and multi-level partitions; and (3) handling the near-boundary ambiguity in model-based spatial clustering techniques. We introduce the extension for obtaining random partitions with more flexibility and less constrains on the shape of clusters comparing to the existing model-based approaches. Moreover, it improves the posterior mixing and captures the variability near the boundary to address the uncertainty. Dimension reduction is achieved in both directions, and we adopt Bayesian wavelet smoothing for the vertical partition. We demonstrate the effectiveness of the proposed method with two applications: (1) the high-dimensional satellite-observed latitude zonal and monthly spectral radiance measurements in climate change study, and (2) the massive functional magnetic resonance imaging data for connectivity analyses.
|
