Text

theses

Forest soil organic carbon (SOC) is the largest terrestrial pool of carbon, and its management plays a significant role in global efforts to mitigate atmospheric carbon concentrations. Despite its importance, much of the world is still lacking good baseline data of forest soil carbon stocks. In the past, broad scale stocks of forest SOC have been derived from soil surveys based on a small number of sampling units, and the resulting estimates are highly uncertain. More recently, predictive statistical models have received attention as an approach for mapping soil carbon at scales relevant to climate change policy and research. However, in order for these models to be useful they must provide full and accurate accounting of uncertainty, in addition to accurate predictions. This dissertation aims to improve prediction of forest SOC by incorporating two potentially important sources of uncertainty into the modeling process: (1) spatial dependence in soil inventory data; and (2) error associated with assuming a single model to be “true”. In order to address these issues, we turn to well established techniques in the Bayesian statistics literature. Our primary focus is on exploring the application of spatial Bayesian hierarchical regression models for improving estimates of forest carbon. This line of research involves both characterizing the spatial dependence in forest SOC inventories at regional, national, and continental scales (the focus of chapters 1 and 3), and applying spatial hierarchical models for mapping SOC and validating this method against non-spatial approaches (chapter 4). Additionally, in chapter 2 we compare methods for model selection and weighting, as well as the effect of model averaging to account for model uncertainty, through rigorous predictive checks. This work is conducted with both forest SOC data as well as other ecological datasets. Taken together, these studies highlight the need for a consistent statistical framework in order to generate reproducible estimates of forest SOC stocks across the globe. Our results argue for hierarchical models, and especially spatial hierarchical models, as a reasonable way forward for predictive mapping of SOC. However, they also highlight significant methodological development that will be necessary in order to obtain predictively accurate models.

ETD_5819

Ph.D.

Includes bibliographical references

by Brian J. Clough

doi:10.7282/T3154FHR

ETD doctoral

The author owns the copyright to this work.