Assessing Model Structure Uncertainty Using Bayesian Model Averaging.

In digital soil mapping and modeling, it is a common practice to ignore model structure uncertainty in the model development process. When there are a range of competing models that are all statistically sound but lead to different predictions or interpretations about the underlying processes, selecting a single model may result in over-confidence in making inference because it underestimates the uncertainty of model structure (form) with regards to the soil properties of interest. Bayesian Model Averaging (BMA) has emerged as a solution to account for the model structure uncertainty and has been successfully used in a variety of disciplines. In this study, the BMA with Markov Chain Monte Carlo (MCMC) samplers was applied to develop predictive models of soil organic carbon (SOC) in Florida, USA. A total number of 1,080 SOC observations were taken at 0-20 cm depth in Florida, USA between 2008 and 2009. To ensure the impartiality in variable selection, an exhaustive set of 210 potential environmental variables (candidate predictors) that represent a comprehensive set of pedogenic and environmental factors were used. Thirty percent of the samples were randomly selected and held out to validate the model developed with the remaining 70 % samples. Results showed that the BMA ensemble prediction was a weighted average of individual predictions and gave higher weights to those ensemble members that were dominated by vegetation and soil moisture predictors. In addition, the BMA ensemble had better predictive accuracy than the ensemble mean and any of the ensemble members. The 95 % prediction intervals were also smaller for the BMA ensembles. The posterior probability density functions (PDFs) from the BMA model also showed a marked amount of uncertainty existing in SOC predictions, especially for high SOC samples, indicating care should be taken when basing decision-making on the predictions from the model.