Gaussian predictive process models for large spatial data sets.

Area of Science:

• Statistical modeling
• Geospatial analysis
• Computational statistics

Background:

• Spatial process models are increasingly used for statistical inference with geocoded data.
• Hierarchical models with Markov chain Monte Carlo (MCMC) are popular but computationally intensive for large datasets.
• Computational complexity, especially with multivariate or spatiotemporal data, limits current methods.

Purpose of the Study:

• To develop computationally efficient predictive process models for large spatial and spatiotemporal datasets.
• To reduce the computational burden associated with traditional hierarchical spatial models.
• To enable flexible modeling of non-stationary, non-Gaussian, multivariate, and spatiotemporal processes.

Main Methods:

• Introduction of predictive process models that project spatial/spatiotemporal processes into lower-dimensional subspaces.
• Utilizing matrix decomposition techniques to reduce computational complexity.
• Development of a computational template applicable to diverse spatial and spatiotemporal settings.

Main Results:

• Predictive process models significantly reduce computational complexity, making large-scale spatial and spatiotemporal data analysis feasible.
• The approach accommodates non-stationary, non-Gaussian, multivariate, and spatiotemporal processes.
• Demonstrated effectiveness through simulations and real-world data analysis.

Conclusions:

• Predictive process models offer a computationally efficient and flexible alternative for analyzing large spatial and spatiotemporal data.
• This methodology expands the scope of statistical inference for complex geospatial datasets.
• The proposed computational template provides a versatile framework for various modeling scenarios.