Digital soil mapping using Random Forest model in Golestan province

Soil maps are the major sources for management of lands, natural resources and environmental aspects. Digital soil mapping methods can decrease the cost of mapping and are not time-consuming. Conditioned Latin Hypercube Sampling for determining the sites of sampling and Random Forest to predict soil sub group classes were evaluated at an area of 85000 ha in Golestan province. 12 environmental variables including terrain attribute, geomorphology units and vegetation cover index were used as predictor parameters. There were11 sub group classes in the soils of study area. The results revealed that the lowest OOB estimate error rate in modeling was 52.53 %. The soil classes with higher frequency had lower OOB error. Gypsic Aquisalids with 22 samples and Aquic Calcixerepts with 9 and Typic Haploxerepts with 22 samples had the OOB error rates 22, 25 and 37%, respectively. Gypsic Haploxerepts (6 samples), Typic Aquisalids (5 samples) Typic Calcixerepts (4 samples), Typic Calcixerolls (4 samples) and Typic Halaquepts (3 samples) had OOB error rate 100%. SAVI, geomorphology, elevation and aspect were the most importance parameters in prediction of soil map and increasing model accuracy. Results indicate that RF technique can be reliable and appropriate method to give satisfactory results with small samples.