From Pedon to Prediction: Innovative Methods for Predicting Soil Distribution On the Landscape.

Predictive (digital) soil mapping involves quantitative relationships between easily measured environmental covariates (slope, aspect, etc.) and field and laboratory observations. In this study soil distribution was predicted for 30,000 hectares of arid upland rangeland with complicated geology in southwestern Utah, USA. Sampling of the environmental covariates was accomplished using Conditioned Latin Hypercube Sampling (cLHS), a stratified random approach which samples the probability distribution of all covariates and returns a sample that closely approximates the original probability distribution of all covariates. Three hundred field locations were identified from the cLHS output and soil and landscape characteristics were described at each site. Data collected for each pedon allowed classification to family and sometimes series level. Prediction of soil distribution was accomplished using Random Forests. Random Forests is a tree based method of classification where multiple uncorrelated trees are grown (a forest) using bootstrap sampling and random subset selection of covariates at each split. Random Forests is non-parametric and typically returns low error rates while providing estimates of variable importance and other useful measures. Numerous levels of soil taxonomy as well as several soil attributes were predicted over the landscape with varying levels of error. Small observation numbers in each predicted class adversely affected classification accuracy.