B1100
Title: Developing spatial multi-resolution models for forestry data with Liesel
Authors: Paul Wiemann - The Ohio State University (United States) [presenting]
Isa Marques - The Ohio State University (United States)
Thomas Kneib - University of Goettingen (Germany)
Abstract: Liesel is a software framework for developing and estimating Bayesian models. Compared to many popular probabilistic programming languages like Stan, Liesel offers full control of the MCMC algorithm. Moreover, when a computationally efficient formulation of the model is available, it can be easily implemented in Liesel. We use these features of Liesel when developing spatial multi-resolution models for data with a specific structure often found in forestry. Here, data is collected intensively in several rather small-sized plots. The plots, however, are distant, and no observations are made between plots. The different intensities in different areas can yield single-resolution stationary spatial models -- these assume the same dependence structure over the whole space -- inappropriate. We present a Bayesian spatial multi-resolution approach that models separately local and global processes. Exploiting the specific structure allows us to model Gaussian random fields with full covariance matrices while keeping the model still computationally feasible. We outline the computational implementation details and compare the performance of the approach presented using simulated and real data.