Dynamic vegetation models can be used to assess the impact of climate change on future vegetation and land cover, including effects on ecosystem services (such and agricultural output) and on carbon dioxide uptake due to vegetation. Understanding the behaviour of dynamic vegetation models, their parametrisation and sensitivity to different climate inputs is of great importance for the accurate assessment of future climate risks. This project aims at addressing both the parametrisation and to analyse the sensitivity of a dynamic vegetation model.
The performance of dynamic vegetation models depends on their parametrization, and matching model output to observations allows us to fine-tune parameters within the model. The resulting optimization problem is multimodal and non-linear, and depending on the degree of non-linearity,
traditional methods based on linearization of model processes might perform poorly. Due to the highly non-linear characteristics of the LPJ-GUESS, we investigate and compare a recently developed stochastic gradient search algorithm against a particle filter method to improve the parametrization of the LPJ-GUESS model. These data assimilation models are both stochastic
methods, which are parallelizable and designed to handle complex characteristics such as multimodality and non-linearity.
For the sensitivity analysis we will focus on predicting future agriculture output under varying climate scenarios. A major problem here is that the vegetation model is computationally expensive to run and to assess all possible future climate scenarios using the vegetation model will be prohibitively expensive. Instead we will try fit a statistic emulator (or approximation) to a suitable subset of model runs. However to fit and evaluate the emulator several runs of the dynamic vegetation model will be required.
The work will be performed by Unn Dahlen (PhD student at the Center for Mathematical Sciences, LU) and a PhD student to be hired during May (also at the Centre). The the project is financed through an eSSENCE and the MERGE strategic research area.