Optimising the Prediction of Soil Properties with Limited Datasets.

With more data available, better soil pit location and faster computers, digital soil mapping (DSM) products are becoming more and more available and their accuracy is improving.  This “revolution” greatly facilitates the use of soil variables into modelling and into the assessment of ecosystem functions or vulnerabilities because maps of soil properties are now becoming available as well as that of the uncertainties of predictions. Predicting soil properties in a forest environment is however generally difficult due to the low density of available soil data and to the limitations in the availability of co-variables. For these reasons it is important to make the best use of the available data.  In the development of most DSM products, only one (or a few) statistical methods have been used and spatial autocorrelation is often not considered.    We tested several statistical methods on the same dataset covering the whole Canadian managed forest. Methods included parametric methods, non-parametric methods of machine learning (Knn, Random forest, boosted regression trees, cubist) and a Bayesian approximation method (INLA), and we tested the effect of including or not a spatial component. The choice of a statistical method had a great impact on the quality of the predictions but the optimal method varied with soil properties, indicating a benefit of testing several approaches. Inclusion of a spatial component greatly improved the predictions in most cases. A possible explanation for this effect is that the processes driving a soil properties, for example soil C, and therefore predictors may vary regionally.