B21G-0142:
Comparing Spatial and Non-Spatial Hierarchical Models for Mapping Forest Soil Organic Carbon at Large Spatial Scales.

Tuesday, 16 December 2014
Brian James Clough, University of Minnesota Twin Cities, Minneapolis, MN, United States and Edwin J. Green, Rutgers University New Brunswick, Ecology, Evolution, and Natural Resources, New Brunswick, NJ, United States
Abstract:
Spatially referenced soil inventory datasets facilitate the mapping of forest soil organic carbon (SOC) at large spatial scales via statistical interpolation. When spatial autocorrelation is present in these data, geostatistical modeling strategies may lead to improved accuracy and a better understanding of the uncertainty within the predictive model. In this study, we compared spatial and non-spatial Bayesian hierarchical models for predicting SOC across forested lands in the nation of Germany. We used observations from the E.U. Joint Research Centre’s LUCAS topsoil database, coupled with predictor variables drawn from remote sensing data products, to address the following objectives: (1) examine patterns of spatial autocorrelation in a national forest SOC dataset; (2) compare spatial and non-spatial models for predicting forest SOC at new locations; and (3) apply the selected model to map predicted soil carbon, along with associated uncertainty estimates, across a grid covering all German forests. Exploratory analyses indicate that there is spatial autocorrelation in the SOC data, and our results suggest that incorporating this spatial dependence within the model framework offers a 9-10 percent reduction in root mean square prediction error (RMSPE) relative to non-spatial models within our study region. By adopting a Bayesian hierarchical approach, where full posterior distributions may be generated at each prediction location, we found significant uncertainty relative to mean estimates when scaling up plot data to the national scale, even when spatial dependence was accounted for. Our results suggest that while accounting for spatial dependence improves predictive performance, difficulty associated with establishing clear relationships between forest SOC and predictor variables limits model precision. By conditioning predictions on both the model parameters and input data, Bayesian hierarchical models were important to our understanding of the model uncertainty. By demonstrating that geostatistical models are a valid approach for mapping forest SOC at broad scales, and that even “best case” predictive methods are likely to be highly uncertain, these results have important implication for efforts to develop soil carbon stock baselines at broad spatial scales.