Parallel cross-validation: A scalable fitting method for Gaussian process models

Abstract Gaussian process (GP) models are widely used to analyze spatially referenced data and predict values at locations without observations. They based on a statistical framework, which enables uncertainty quantification of the model structure predictions. Both evaluation likelihood prediction involve solving linear systems. Hence, computational costs large limit amount that can be handled. While there many approximation strategies lower cost GP models, they often provide sub-optimal support for parallel computing capabilities (high-performance) environments. To bridge this gap parallelizable parameter estimation method is presented. The key idea divide spatial domain into overlapping subsets use cross-validation (CV) estimate covariance parameters in parallel. Although simulations show CV less effective than maximum method, it amenable handling datasets. Exploiting screen effect helps arrive analysis close global computation despite performing computations local regions. Simulation studies assess accuracy estimates implementation shows good weak strong scaling properties. For illustration, an exponential fitted scientifically relevant canopy height dataset with 5 million Using 512 processor cores brings time one configuration 1.5 minutes.