Evaluation of Accuracy of Digital Soil Mapping with Limited Data in a Part of Loess Plateau, Golestan Province

 Digital soil mapping approaches that require quantitative data for prediction are difficult to implement in countries with limited data on soil and auxiliary variables. Extensive field sampling is very labor intensive and costly that is problematic for mapping missions. On the other side, it is believed in digital soil mapping approaches that unique soil conditions (soil types or soil properties) can be associated with unique combination and configuration of environmental variables. In this study we used a Random Forest (RF) algorithm combined with classification information of 64 soil profiles and 19 environmental variables (including terrain attribute, geomorphology units, land use and vegetation cover index) to map soil classes in the part of loess plateau, Golestan Province Iran. Geomorphology, elevation, slope aspect and land use were the most important parameters in prediction of soil map in different taxonomic level. The results of accuracy assessment of RF with different entrance variables revealed that accuracy of model including overall accuracy and kappa index respectively decreased of 0.91 and 0.83 for great group, 0.78 and 0.56 for subgroup, 0.50 and 0.32 for family. The minimum and maximum Out of bag (OOB) estimate error rate in modeling were 32.69% and 65.38% for great group and soil family, respectively and the soil classes with higher frequency had the lower OOB error. The present study showed that in regions of Iran with limited data, digital soil mapping and high resolution ancillary data with smaller sample size can be led to an effective result in higher taxonomic levels.