High-Dimensional Bayesian Geostatistics.

Area of Science:

• Statistics
• Geographic Information Systems (GIS)
• Computational Statistics

Background:

• Hierarchical spatiotemporal models are crucial for understanding spatial and temporal variability.
• Traditional models face computational intractability with large datasets due to cubic complexity.
• Scalability is a major challenge in analyzing big spatiotemporal data.

Purpose of the Study:

• To review two methods for constructing highly scalable spatiotemporal stochastic processes.
• To provide model-based solutions for big spatiotemporal datasets.
• To enable efficient Bayesian inference for complex spatiotemporal fields.

Main Methods:

• Low-rank processes operating on lower-dimensional subspaces.
• Nearest-Neighbor Gaussian Process (NNGP) ensuring sparse precision matrices.
• Embedding scalable priors within hierarchical modeling frameworks.

Main Results:

• Both methods offer algorithmic complexity of approximately n floating point operations (flops) per iteration.
• These scalable processes can be used as priors for spatiotemporal random fields.
• Facilitates full Bayesian inference for large-scale spatiotemporal data.

Conclusions:

• The reviewed methods provide computationally efficient solutions for big spatiotemporal data analysis.
• These approaches enhance the feasibility of using complex hierarchical models for large datasets.
• Offers a pathway to better understand spatial and temporal variability in big data contexts.