Modelling and Prediction of Soil Classes Using Boosting Regression Tree and Random Forests Machine Learning Algorithms in Some Part of Qazvin Plain

Appropriate selection of ancillary covariates have a specific important on digital soil mapping. Currently, use of machine learning algorithms for digital mapping and updating of conventional soil map has been developed in Iran. The current study has been done to compare the BRT and RF models for spatial prediction of subgroup and family classes with selection of axillary variables  using VIF approach in some part of Qazvin Plain. 61 pedons were sampled based on stratified random, digged, described and classified with consideration of laboratory analysis up to family level. The most appropriate variables were selected among 15 Geomorphometry and Remote Sensing Indices using Variance Inflation Factor (VIF). Soil landscape modeling was conducted with RF and BRT learning algorithm in RStudio software based on Randomforest and C5.0 packages at subgroup and family levels. The results showed that six indices including CHA, DEM, STH, SI DVI and NDVI were selected as input variables. Assessment indices such as the Overall Accuracy (OA) and Kappa were obtained for BRT (35, 26%) and RF (70, 60%) at family level, respectively. Sensitivity analysis based on the mean decrease accuracy (MDA) revealed that the modified catchment area variable is the most relative important variable among the selected variables. Generally, by using feature selection innovative approach and effective learning algorithms, the spatial distribution of soil maps could be made even in low relief lands with acceptable accuracy.