Predicting soil organic carbon percentage from loss-on-ignition using Bayesian Model Averaging

Abstract

A shortage of data for percentage of organic carbon (C%) makes calculation of soil profile carbon storage difficult. Loss on ignition (LOI) data, which are cheap to obtain and often readily available, can be used to estimate organic C%. This paper simultaneously considers several predictors of organic C%: LOI, parent material, drainage status, type of soil horizon, clay content, and pH. In order to model appropriately the existence of multiple hypotheses and the consequent model uncertainty, a Bayesian Model Averaging (BMA) approach was used. BMA considers all models that result from all possible combinations of explanatory variables. Based on a BMA approach and Scottish Soil Survey data, it was found that the most important factors to predict organic C% were LOI, clay content, a dummy for Countesswells Association (till derived from granite), and a dummy for B horizon soils. The validation analysis showed that prediction accuracy for organic C% was better with the BMA approach than with an ordinary least-squares approach that includes no other predictors apart from LOI (i.e. 22% reduction in horizons A, Ap, and C).