Soil Distribution Pattern Analysis in a Low Relief Area Using Decision Trees Algorithm

Digital soil mapping (DSM) can be defined as a production of spatial soil information. Decision tree (DT) algorithm is one of the most popular machine learning methods which was applied in several recent DSM studies. This study was carried out to evaluate the capability of DT in mapping soils in Miandarband region with area of 50,000 ha in Kermanshah province. The C5.0 decision tree algorithm (with and without boosting meta-algorithm) used to establish spatial relationships between known soil taxonomic classes and environmental variables. Using simple systematic sampling, 78 pedons were studied and 6 great groups and 14 subgroups of Soil Taxonomy (ST) were identified. Thirty environmental items were derived from a digital elevation model (DEM) file and a landsat-8 OLI/TIRS (July/Tir 1394) image of the area. Predictions made by C5.0 algorithm showed OA values of 73 percent for great group and subgroup, while comparable values for Kappa Index were 0.61 and 0.63, respectively. Combination of boosting meta-algorithm with C5.0 increased OA values for ST categories 0.80 and 0.76 and Kappa Index values to 72 percent and 66 percent. Results showed a considerable capability for DT in recognition of soil pattern over the study area and the topographic variables seems to be most important. Also, analysis of the produced maps, compared with the observed soil pattern during the field survey, revealed a reasonable agreement of decision tree algorithm predictions with reality.