QUANTIFICATION AND INCORPORATION OF UNCERTAINTY IN FOREST GROWTH AND YIELD PROJECTIONS USING A BAYESIAN PROBABILISTIC FRAMEWORK (A DEMONSTRATION FOR PLANTATION COASTAL DOUGLAS-FIR IN THE PACIFIC NORTHWEST, USA)

A Bayesian probabilistic modeling platform was used and evaluated for application in a relatively complex individual-tree growth and yield model for coastal Douglas-fir (Pseudotsuga menziesii var. menziesii (Mirb.) Franco), which was expressed as a mixed discrete and continuous Bayesian Network for annual projections. The modeling platform used a common and open-source Bayesian analysis program (JAGS v3.3.0), and was sufficiently flexible to handle a relatively complex model structure; namely, a differential form, highly dynamic, recursive, hierarchical, non-linear system of equations with rather complex error structures. This novel probabilistic modeling platform met certain desirable criteria, including: (1) accurate and tractable projections that included full error propagation; (2) flexible and comprehensive analytic capabilities; (3) full consideration of hierarchical and multi-level model structures; (4) capacity for random effects calibration; (5) allowance of hypothesis testing and updating knowledge across different system components, simultaneously with varying sources of information (i.e., new data); (6) computational efficiency; and (7) relatively simple implementation as demonstrated in a compiled scripting language. Probabilistic projections of forest growth and yield included all sources of errors and uncertainty (e.g., estimated parameters, state variables, random effects, and residual errors). Cumulative error projections over a 40-year period for three sample Douglas-fir stands were determined. Projection errors for key metrics summed across all trees, such as total basal area and stem density, had coefficient of variations between 4-6% and 7-8%, respectively. Probabilistic projections were markedly different from deterministic projections made with the same model structure. Overall, this novel probabilistic platform showed strong promise as a general platform for ecological modeling, particularly when tractable and analytically correct error projections are required. In particular, the Bayesian probabilistic modeling approach used provided a natural platform for cross-disciplinary research, particularly between social and ecological research domains.