Abstract:
|
Gaussian process models have been widely used and accepted as the tool of choice in spatial statistics with continuous outputs. However, these models face tremendous modeling and computational challenges for very large heterogeneous spatial data sets. The ability to model the correlation and heterogeneity of the data with accuracy is of particular importance in environmental and geophysical sciences. It provides better prediction and useful information for both the global and the local scale of the data. To address these challenges, we propose a new non-separable predictive process Gaussian process model. We also induce nonstationarity in the model with a Bayesian adaptively selected partitions.
|