Incorporating taxonomic distance into spatial prediction and digital mapping of soil classes

Abstract Mapping soil classes digitally generally starts with soil profile description with observed soil classes at a taxonomic level in a particular classification system. At each soil observation location there is a set of co-located environmental variables, and the challenge is to correlate the soil class with the environmental variables. The current methodology treats soil classes as ‘labels’ and their prediction only considers the minimisation of the misclassification error. Soil classes at any taxonomic level have taxonomic relationships between each other, and in some instances the errors in prediction of certain classes are more serious than the others. No statistical procedure so far has been utilised to account for these relationships. This paper shows that in digital mapping of soil classes, we can incorporate the taxonomic distance between soil classes in a supervised classification routine. Using classification trees, we can specify an algorithm that minimises the taxonomic distance rather than misclassification error. Two examples are given in this paper for mapping soil orders in the Australian soil classification system. A site in the Edgeroi area showed the advantage of using the method that minimises the taxonomic distance. Meanwhile a site in the Hunter Valley showed minimising the misclassification error performed similarly to minimising taxonomic distance. The advantages and challenges of using soil taxonomic distance are discussed.