Semi-automated disaggregation of a conventional soil map using knowledge driven data mining and random forests in the Sonoran desert, USA

Conventional soil maps (CSM) have provided baseline soil information for land use planning for over 100 years. Although CSM have been widely used, they are not suitable to meet growing demands for high resolution soil information at field scales. We present a repeatable method to disaggregate CSM data into ~30-meter resolution rasterized soil class maps that include continuous representation of probabilistic map uncertainty. Methods include training set creation for original CSM component soil classes from soil-landscape descriptions within the original survey database. Training sets are used to build a random forest predictive model utilizing environmental covariate maps derived from ASTER satellite imagery and the USGS National Elevation Dataset. Results showed agreement at 70 percent of independent field validation sites and equivalent accuracy between original CSM map units and the disaggregated map. Uncertainty of predictions was mapped by relating prediction frequencies of the random forest model and success rates at validation sites.