Parameterization of a process-based tree-growth model: Comparison of optimization, MCMC and Particle Filtering algorithms

Abstract

Finely tuned process-based tree-growth models are of considerable help in understanding the variations of biomass increments measured in the dendrochronological series. Using site and species parameters, as well as daily climate variables, the MAIDEN model computes the water balance at ecosystem level and the daily increment of carbon storage in the stem through photosynthesis processes to reproduce the structure of the tree-ring series. In this paper, we use three techniques to calibrate this model with Pinus halepensis data sampled in the Mediterranean part of France: a standard optimization (PEST), Monte Carlo Markov Chains (MCMC) and Particle Filtering (PF). Contrary to PEST, which tries to find an optimum fit (giving the lowest error between observations and simulations), the principle of MCMC and PF is to walk, from a priori distributions, in the parameter space according to particular statistical rules to compute each parameter distribution. The PEST and MCMC calibrations of our dendrochronological series lead to rather similar adjustments between simulations and observations. PF and MCMC calibrations give different parameter distributions, showing how complementary are these methods, with a better fit for MCMC. Yet, independent validations over 11 independent meteorological years show a higher efficiency of the recent PF method over the others.