Updating soil survey maps using random forest and conditioned Latin hypercube sampling in the loess derived soils of northern Iran

Abstract Many Iranian soil surveys need to be updated. Conventional soil survey methods are expensive and time-consuming. Digital soil mapping (DSM) can be used for updating soil surveys. Many sampling and modeling techniques exist for DSM. In this paper we investigate the use of conditioned Latin hypercube sampling and random forest modeling for mapping Soil Taxonomy great group, subgroup and series levels for ~ 85,000 ha in Golestan Province, Iran. Overall error was 48.5, 51.5 and 56.6% for great group, subgroup and series levels, respectively. Estimated individual soil type error was between 8 and 100%. Soil types with larger sample sizes were predicted over a greater area at each taxonomic level. The soil adjusted vegetation index, the conventional soil series map and geomorphology were the most important covariates for each taxonomic level. Taxonomic classes with important covariates had low OOB error. The updated soil series map was 13.4% more accurate than the existing conventional soil series map.