The application of classification tree analysis to soil type prediction in a desert landscape

Classification tree analysis is evaluated as a predictive soil mapping technique for developing a preliminary soil map for neighboring site from samples extracted from an existing soil map. The objective of the research is to help guide future soil mapping in a nearby area. In order to determine the best overall modeling approach several variations were explored: the dependent variable (soil map class) was grouped at several hierarchical levels (according to Soil Taxonomy), sensitivity analysis was performed on the predictor variables (environmental variables acting as surrogates for soil forming factors), and the study area was divided into meaningful sub-areas (mountains and basins). Soil great group was discovered the most parsimonious dependent variable based on model results (misclassification error rate of 30.0% based on a test data set). Geomorphology (as measured by several landform variables) best explains the distribution of soil types. The terrain analysis variables did not explain a large amount of variance within the models. Dividing the study area in two separate modeling units increased overall model accuracy. Our results suggest that soil taxonomic class can be predicted with reasonable accuracy from environmental variables. In addition, the technique can provide limited insight into the variables that are most responsible for driving soil development in a given area. This technique could be used in soil survey to extrapolate obvious soil landscape relationships from one site to another, allowing soil experts to concentrate their field mapping effort in unique areas.