Bayesian methods applied for ecosystem model calibration and uncertainty source estimation

Abstract

How significant are different uncertainty sources when simulating the future state of the ecosystem in Finland? In this thesis, we examine this question and provide some answers to this broad topic by simulating 21 st century ecosystem conditions with a land-ecosystem model called JSBACH. The results are also compared to similar simulations performed by another model called PREBAS. We consider four different sources of uncertainty that are related to 1) the model that is used to generate the future conditions; 2) future climate used to drive the model, represented by an ensemble of CMIP5 simulations; 3) RCP scenarios that depict the rising atmospheric CO2 concentration and; 4) forest management actions. Before running the simulations described above, we calibrated and validated the JSBACH model extensively on different temporal resolutions and with multiple model modifications. These hindcasting calibrations were performed with two Bayesian approaches: the adaptive Metropolis algorithm and the adaptive population importance sampler. The calibrations resulted in a sufficient model setup and satisfactory parameter distributions. These were used to represent the JSBACH model uncertainty in the 21 st century simulations. Canonical correlation analysis was used to gleam the impact of the different uncertainty sources on multiple groups of ecosystem variables. The results are summarised via the use of redundancy indices that yield varied impacts. Overall, forest management actions and RCP scenarios tend to dominate the uncertainties towards the end of the century, but the effect of climate models and parameters should not be overlooked especially since a more detailed examination revealed that their impact was not fully captured. Publishing unit Classification (UDC)

Potential of Bayesian formalism for the fusion and assimilation of sequential forestry data in time and space

“…Regardless, the method exhibited equal robustness compared to other state-of-the-art methods for forest inventories. Similarly, Mäkelä (2020) addressed the uncertainty in ecosystem modeling through the Bayesian approach of canonical correlation analysis (CCA; Hotelling, 1937), which is a technique for detecting correlations between two multivariate or random variables and extracting linear components that represent the correlation. The method was equally useful in identifying the uncertainty caused by varied factors on ecosystem modeling.…”

“…For instance, it found that forest management was the dominant factor that contributed to the uncertainty in the study. The idea of identifying uncertainty elements is intriguing, especially with Bayes, as illustrated by Varvia (2018) and Mäkelä (2020).…”