Applications of Bayesian computational statistics and modeling to large-scale geoscientific problems

Abstract

Climate change is one of the most important, pressing, and furthest reaching global challenges that humanity faces in the 21st century. Already affecting daily lives of many directly and everyone indirectly, changes in climate are projected to have many catastrophic consequences. For this reason, researching climate and climate change is needed.

Studying complex geoscientific phenomena such as climate change consists of a patchwork of challenging mathematical, statistical, and computational problems. To solve these problems, local and global process models and statistical models are combined with both small in situ observation data sets with only few observations, and equally well with enormous global remote sensing data products containing hundreds of millions of data points. This integration of models and data can be done in a Bayesian inverse modeling setting if the algorithms and computational methods used are chosen and implemented carefully. The methods used in the four publications on which this thesis is based range from high-dimensional Bayesian spatial statistical models and Markov chain Monte Carlo methods to time series modeling and point estimation via optimization.

The particular geoscientific problems considered are: finding the spatio-temporal distribution of atmospheric carbon dioxide based on sparse remote sensing data, quantifying uncertainties in modeling methane emissions from boreal wetlands, analyzing and quantifying the effect of climate change on growing season in the boreal region, and using statistical methods to calibrate a terrestrial ecosystem model.

In addition to analyzing these problems, the research and the results help to understand model performance and how modeling uncertainties in very large computational problems can be approached, also providing algorithm implementations on top of which future efforts may be built.