In most of the arid and semi-arid regions in the world, soil salinization is one of the main serious problems for agriculture and sustainable land use management. This study assessed the applicability of digital soil mapping (DSM) using a combination of terrain attributes and remote sensing data to predict and map soil salinity in Kurdistan Province, Iran. Using the randomized sampling method, 150 soil locations were chosen from the study area, and then from surface layers (0-30 cm) and subsurface layers (30-60 cm) soil samples were collected. First, using Boruta algorithm the most relevant environmental covariates were selected. Then, the relationship between EC contents and the selected environmental covariates were constructed using Random Forest (RF), Support Vector Machine (SVM), and Cubist models. Also, 10-fold cross validation method was used to evaluate models. Results showed that the increment of EC with soil depth increment. The highest performance (RMSE and R2) for prediction of EC in two depths were achieved using the RF model. More specifically, the R2 of prediction was up to 3% and 12% better in comparison to Cubist and SVM, respectively. Moreover, our results showed that an improvement was obtained in the prediction accuracy of models with using Boruta algorithm to select the most relevant covariates. As a result, Boruta-based Random Forest model was recommended for mapping soil salinity using environmental covariates derived from DEM and satellite in other semi-arid regions.
