|
Abstract:
|
In 2013, the U.S. Geological Survey completed a geochemical and mineralogical soil survey of the conterminous United States. At 4,857 locations, soil samples were collected at three depth levels; for each sample, the concentrations of 45 elements and 20 minerals were measured. To aid interpretation of these multivariate, compositional data sets, statistical clustering was used. The clustering procedure partitioned the field samples for a data set into two clusters. Each cluster was partitioned again to create two sub-clusters, and so on, generating a hierarchy of clusters. The different levels of the hierarchy indicated different influences of the soil-forming processes at varying spatial scales. The clustering method was based on a Bayesian finite mixture model. The model parameters were estimated with Hamiltonian Monte Carlo sampling of the posterior probability density function, which usually had multiple modes. Each mode was associated with a unique set of model parameters. The set that was most consistent with the independent geologic knowledge was selected for detailed interpretation and further partitioning of the data.
|
