Modeling Spatio-Temporal Dynamics of the High Plains Aquifer Using a Dimension-Reducing Nearest-Neighbor Gaussian Process

• Activity Number: 116
• Type: Topic Contributed
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Andrew Oliver Finley*+ and Sudipto Banerjee and Abhirup Datta
• Companies: Michigan State University and University of Minnesota and University of Minnesota
• Keywords: Nearest-Neighbor Gaussian process ; Bayesian hierarchical ; water

Abstract

This talk focuses on predicting water-level dynamics across the High Plains Aquifer. This aquifer supplies 30 percent of the US irrigated groundwater and extends beneath parts of eight states in the Great Plains. Measurements of depth to water table have been collected from agricultural irrigation wells across the Aquifer since the late nineteenth century resulting in 1,299,892 measurements. The proposed dynamic space-time water-level model uses a new class of dimension-reducing spatial models based upon a well-defined Nearest-Neighbor Gaussian process (NNGP). We demonstrate how the model can be embedded within a versatile Bayesian hierarchical modeling framework to offer full Bayesian inference using a matrix-free Gibbs sampling algorithm. The model's scalability to massive datasets such as those found in climate sciences far exceeds those of process-based low rank models such as the predictive process. Importantly, we show the NNGP does not suffer from the inferential limitations of low-rank models and is able to effectively reproduce the corresponding inference from full (but highly expensive) geostatistical models.