Two Methods for Using Legacy Data in Digital Soil Mapping

Legacy data are useful sources of information on the spatial variation of soil properties. There are, however, problems using legacy data, and in this paper we explore some of these problems. A common issue is often the uneven sample distribution over geographical and predictor space and the problems this generates for the subsequent modelling efforts. Furthermore legacy soil data often has a mixture qualitative and quantitative data. The current need is for quantitative data, whereas the available datasets are often qualitative; e.g. auger bores. In this paper we compare two methods: (i) a Generalized Linear modelling (GZLM) approach which uses scarce, measured soil property data and (ii) Bayesian Belief networks (BBN) which uses extensive but generic values of the soil property, linked to soil classes. We used digital soil mapping covariates such as small scale soil maps, geology, digital terrain model, climate data and landscape position in order to predict continuous surfaces for sand, silt, clay, bulk density and organic carbon. The objective is to present a qualitative comparison between the two methods, as a direct comparison was not possible due to the number and distribution of the legacy data. We found that the GZLM approach was significantly impacted by an uneven sampling of the predictor space. This study suggests that a more generalist approach such as BBN is better in the absence of few hard data but in the presence of many soft data.