Computational Considerations for Applying Nearest Neighbor Gaussian Processes to Large Spatial Data Sets: A Case Study from Forest Biomass Prediction Across Alaska

• Activity Number: 589
• Type: Invited
• Date/Time: Wednesday, August 3, 2016 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Sudipto Banerjee and Abhirup Datta and Bruce Cook and Andrew Finley*
• Companies: University of California at Los Angeles and University of Minnesota and NASA and Michigan State University
• Keywords: Gaussian Process ; MCMC ; Forestry

Abstract

Monitoring US forest carbon stocks is critical for natural resource management and greenhouse gas reporting activities. This talk will cover the modeling advancements that made possible the first regional estimates of forest carbon stocks for the Tanana Inventory Unit of interior Alaska (146,000 km2). The proposed multivariate hierarchical spatial model addresses several data challenges including spatial misalignment and change-of-support among the available: i) sparse network of forest inventory plots; 2) limited extent high spatial resolution airborne data collection with Goddard's LiDAR, Hyperspectral, and; 3) complete coverage, but spatially coarse spaceborne Landsat imagery. Due to the large amount of data, inference at the desired spatial resolution made it impossible to estimate key spatial process parameters using traditional methods. Rather, we apply a highly scalable Nearest Neighbor Gaussian Process (NNGP) to provide fully model-based inference. We embed the NNGP as a sparsity-inducing prior within the proposed hierarchical modeling framework and outline how computationally efficient MCMC algorithms are executed to deliver the desired forest biomass data products.