OOS 50-10 - Computationally efficient, large-scale application of individual- and trait-based models of tree growth and mortality

Friday, August 11, 2017: 11:10 AM
Portland Blrm 256, Oregon Convention Center
Kiona Ogle1, Jarrett Barber1, Michael Fell1,2, Adam Leighton3 and Jeremy W. Lichstein4, (1)School of Informatics, Computing, and Cyber Systems, Northern Arizona University, Flagstaff, AZ, (2)School of Life Sciences, Arizona State University, Tempe, AZ, (3)School of Mathematical and Statistical Sciences, Arizona State University, (4)Department of Biology, University of Florida, Gainesville, FL

Individual-based models (IBMs) of tree growth that account for species differences have the potential to greatly improve predictions of forest dynamics. Yet, such IBMs are impractical to apply at scales that represent millions of trees. We discuss two approaches to confronting such an IBM with forest inventory data; here, IBM inputs include 32 physiological, anatomical, and allometric traits (parameters), and outputs include annual tree-level heights, stem diameters, and carbon/biomass pools. The first approach aims to understand the trait space of trees by fitting the IBM to USFS Forest Inventory and Analysis (FIA) data for 1.6M healthy, growing trees. Here, we address computational issues by aggregating the data; repeated observations of diameters and heights were summarized across all trees as a multi-dimensional probability array, which we compared against IBM outputs to eliminate “unrealistic” trait values. The second approach preserves the individual-level data, but this negates applying the IBM to millions of trees. To overcome this hurdle, we develop a computationally fast emulator as a prior model for the IBM. We update the emulator using a computationally affordable number of runs from the IBM, giving refined posterior estimates of traits (parameters) and enabling predictions of tree growth at broad scales.


Previously, we fit the IBM to a small FIA dataset within a fully Bayesian framework that employed Markov chain Monte Carlo (MCMC) methods. If such an approach were applied to the 1.6M FIA observations, the MCMC procedure would take ~25 years to complete (without parallelization). Our first approach that fit the IBM to aggregated data via a simple MCMC routine took ~17 hours. This approach produced a complex, multi-dimensional trait space (posterior) that represents the range of possible trait values and correlations associated with realistic tree growth. In particular, tradeoffs between pairs of traits are governed by other traits, and interactions among traits explained 80% of the variation in radiation-use efficiency and ~40% of the variation in a subset of traits related to crown/stem allometries and maintenance respiration. Half of the traits appeared to vary independently of all other traits. While this approach provided valuable insight into the trait space of healthy North American trees, it is limited in its ability to quantify how traits may vary across sites and within and among species. Thus, we compare results from our aggregated data approach to those from our emulator, informed by individual tree data.