Estimation and potential improvement of the quality of legacy soil samples for digital soil mapping

Abstract

Legacy soil data form an important resource for digital soil mapping and are essential for calibration of models for predicting soil properties from environmental variables. Such data arise from traditional soil survey. Methods of soil survey are generally empirical and based on the mental development of the surveyor, correlating soil with underlying geology, landforms, vegetation and air-photo interpretation. There are no statistical criteria for traditional soil sampling, and this may lead to biases in the areas being sampled. The challenge is to test the use of legacy data for large-area mapping (e.g. national or continental extents) in order to limit the funds of field survey for large-area mapping. The problem is then to assess the reliability and quality of the legacy soil databases that have been mainly populated by traditional soil survey, and if there is a possibility of additional funding for sampling, to determine where new sampling units should be located. This additional sampling can be used to improve and validate the prediction model. Latin hypercube sampling (LHS) has been proposed as a sampling design for digital soil mapping when there is no prior sample. We use the principle of hypercube sampling to assess the quality of existing soil data and guide us to locations that need to be sampled. First an area is defined and the empirical environmental data layers or covariates are identified on a regular grid. The existing soil data are matched with the environmental variables. The HELS algorithm is used to check the occupancy of the legacy sampling units in the hypercube of the quantiles of the covarying environmental data. This is to determine whether legacy soil survey data occupy the hypercube uniformly or if there is over- or under-observation in the partitions of the hypercube. It also allows posterior estimation of the apparent probability of sample units being surveyed. From this information we can design further sampling. The methods are illustrated using legacy soil samples from Edgeroi, New South Wales, Australia, and from a large part of the Danube Basin. One third of the total number of sampling units are added to the original dataset. These new sampling units improve the representation of the feature space of the covariate. The standard deviation of the overall density is consequently smaller.