Bayesian calibration of a carbon balance model PREBAS using data from permanent growth experiments and national forest inventory

Abstract

Policy-relevant forest models must be environment and management sensitive and provide unbiased estimates of predicted variables over their intended areas of application. While empirical models derive their structure and parameters from representative data sets, process-based model (PBM) parameters should be evaluated in ranges that have a biological meaning independently of output data. At the same time PBMs should be calibrated against observations in order to obtain unbiased estimates and an understanding of their predictive capability. By means of model data assimilation, we Bayesian calibrated a forest model (PREBAS) using an extensive dataset that covered a wide range of climatic conditions, species composition and management practices. PREBAS was calibrated for three species in Finland: Scots pine (Pinus sylvestris L.), Norway spruce (Picea abies [L.] H. Karst.) and Silver birch (Betula pendula L.). Data assimilation was strongly effective in reducing the uncertainty of PREBAS parameters and predictions. A country-generic calibration showed robust performances in predicting forest variables and the results were consistent with yield tables and national forest statistics. The posterior predictive uncertainty of the model was mainly influenced by the uncertainty of the structural and measurement error.