BAYESIAN OPTIMIZATION OF A DYNAMIC VEGETATION MODEL

Sammanfattning

In this thesis, we apply Bayesian Optimization to estimate the parameters of the LPJ-GUESS dynamic vegetation model with a focus on Methane (CH4) emissions in the Siikaneva wetland in Finland. Previous research has used other statistical methods, such as Markov Chain Monte Carlo (MCMC), to estimate the LPJ-GUESS model parameters, achieving a root mean square error (RMSE) of 0.023. Our objective is to assess whether Bayesian Optimization using a Gaussian Process can yield more accurate parameter estimates with lower error, using fewer computer resources. To evaluate this, we first conduct a twin experiment using synthetic data, testing different acquisition, covariance, and loss functions in the Gaussian Process. We found that the Log Expected Improvement (Log-EI) acquisition function offered a good balance between accuracy and computational cost. The Matérn covariance with ν = 2.5 was found to be suitable for the optimization, and RMSE proved to be more effective than mean absolute error (MAE) as a loss function. We initially assumed a zero-mean Gaussian Process, but then incorporated a mean structure. When applying the optimization to real observations of CH4 emissions, Bayesian Optimization improved parameter estimates and reduced the RMSE to 0.013 — a better result than previously obtained when using MCMC. These results suggest that Bayesian Optimization is an effective approach for parameter estimation in the LPJ-GUESS model.