NONSEPARABLE DYNAMIC NEAREST NEIGHBOR GAUSSIAN PROCESS MODELS FOR LARGE SPATIO-TEMPORAL DATA WITH AN APPLICATION TO...

Area of Science:

• Environmental Science
• Statistical Modeling
• Public Health

Background:

• Particulate matter (PM) poses significant risks to human health, necessitating accurate spatial-temporal monitoring.
• High-resolution maps are crucial for regulatory bodies to identify areas exceeding PM concentration limits.
• Traditional Gaussian Process (GP) models struggle with computational demands of large-scale spatio-temporal data.

Purpose of the Study:

• To develop a scalable computational framework for analyzing large spatio-temporal air quality datasets.
• To introduce a novel Dynamic Nearest Neighbor Gaussian Process (DNNGP) model for efficient PM level prediction.
• To enhance the accuracy of PM level predictions for regulatory and public health applications.

Main Methods:

• Construction of a novel class of scalable Dynamic Nearest Neighbor Gaussian Process (DNNGP) models.
• Utilizing DNNGP as a sparsity-inducing prior for spatio-temporal random effects in Bayesian hierarchical models.
• Applying the DNNGP model to a large-scale European air quality dataset, integrated with LOTOS-EUROS chemistry transport models (CTMs).

Main Results:

• The DNNGP model demonstrates linear storage and memory requirements, enabling massive scalability.
• DNNGP provides substantially superior approximations of spatio-temporal processes compared to low-rank methods.
• The model significantly improved predictions of particulate matter levels across Europe.

Conclusions:

• The developed DNNGP models offer a computationally efficient and statistically robust solution for spatio-temporal air quality modeling.
• This approach overcomes the limitations of traditional GP models, enabling high-resolution mapping of environmental pollutants.
• The enhanced prediction accuracy supports regulatory efforts and public health initiatives by identifying high-risk areas for particulate matter exposure.