Approximate Bayesian computation to recalibrate individual-based models with population data: Illustration with a forest simulation model

Abstract

Listen

Translate article

Ecology makes an increasing use of complex simulation models. As more processes and model parameters are added, a comprehensive model calibration with process-level data becomes costly and predictions of such complex models are therefore often restricted to local applications. In this context, inverse modelling techniques enable to calibrate models with data of the same type than model outputs (thereafter called population data for the sake of clarity, although other data types can be used according to model outputs), which are usually simpler to collect and more readily available. This study aims at demonstrating how such data can be used to improve ecological models, by recalibrating the most influential parameters of a complex model in a Bayesian framework, and at providing general guidelines for potential users of this approach. We used the individual-based and spatially explicit forest dynamics simulation model Samsara2 as a case study. Considering the results of an initial calibration and of a sensitivity analysis as prerequisites, we assessed whether we could use approximate Bayesian computation (ABC) to recalibrate a subset of parameters on historical management data collected in forests with various ecological conditions. We propose guidelines to answer three questions that potential users of the approach will encounter: (1) How many and which parameters are we able to recalibrate accurately with such low-informative data? (2) How many ABC simulations are required to obtain a reasonable convergence of the parameter posterior estimates? (3) What is the variability of model predictions following the recalibration? In our case study, we found that two parameters by species could be recalibrated with forest management data and that a relatively low number of simulations (20,000) was sufficient. We finally pointed out that the variability of model predictions was largely due to model stochasticity, and much less to ABC recalibration and initial calibration uncertainties. Combining direct process-level calibration to ABC recalibration of the most influential parameters opens the door to interesting modelling improvements, such as the calibration of forest dynamics along environmental gradients. This general approach should thus help improve both accuracy and generality of model-based ecological predictions.