Generalization of digital soil mapping results for prediction of soil classes (A Case Study: Shahrekord Plain, Chaharmahal-Va-Bakhtiari Province)

Digital soil mapping (DSM) predicts the soil variability based on the relationship between the soil classes and auxiliary information. Therefore, it is expected that if two regions have similar auxiliary information, the model developed to estimate soil variability for one of these regions could be generalized to the other. The aim of this study was to predict the soil classes up to great group level of Soil Taxonomy (ST) and Reference Soil Group (RSG) level of World Reference Base for the Soil Resources (WRB) across the area with little soil data (recipient site) on the basis of constructed model in area that has sufficient soil data (reference site) using DSM approaches in the Shahrekord plain of Chaharmahal-Va-Bakhtiari Province. The reference and recipient sites are located in the Shahrekord region of Chaharmahal-Va-Bakhtiari Province. The Mahalanobis distance is used to determine the distance between the mean of the reference’s soil forming factors and the recipient’s soil forming factors (Mallavan et al. 2010). The reference site for this study was surveyed using digital soil mapping approaches at semi-detailed scale (i.e., raster maps with pixel size 50×50 m) up to family level by Mosleh (2016). Different machine learning algorithms consisting of artificial neural networks (ANNs), boosted regression tree (BRT), random forest (RF) and multinomial logistic regression (MLR) were considered for each soil taxonomic level to identify the relationship between soil classes and auxiliary information. Fifteen pedons were excavated at the recipient site with 750 m intervals. All the pedons were described and the soil samples were taken from different genetic horizons, air dried, crashed and passed through a 2 mm sieve. The soil samples were classified the soils according to the Soil Taxonomy (Soil Survey Staff 2014) and the WRB (IUSS Working Group WRB 2015) up to great group and Reference Soil Group levels, respectively. The results showed that the Mahalanobis distance at the reference and recipient sites is equal. Therefore, the two studied sites are entirely similar and can be considered as Homosoil. Summary statistics of auxiliary information for the reference and recipient sites indicated that the difference between the mean of the reference’s soil forming factors and the recipient’s soil forming factors is negligible. Extrapolated models across the recipient site lead to similar results with the reference site. These results include: (i) no significant differences were observed between different models to predict soil classes based on the ST system; (ii) OA values showed a decreasing trend with increasing the taxonomic levels for all the studied models (Figure 3); (iii) the MLR model has the highest performance to predict the RSG. The results indicated that DSM could be used for prediction of the soil classes in the Homosoil framework (both sites have similar auxiliary information or soil forming factors). It is expected that the accuracy of predictions is accrued if there is a high agreement between the reference and the recipient sites in terms of the auxiliary information.