Digital mapping of soil classes using spatial extrapolation with imbalanced data

Digital mapping of soil classes using the extrapolation approach is timesaving, economically cheap, and helps collecting soil data from areas with difficult sampling. However, it has not been explored widely enough for digital mapping of soil classes. This study seeks to evaluate and compare several machine learning and regression algorithms for the extrapolation of soil sub-groups. Also the issue of imbalanced number of observations was addressed and oversampling technique was applied on the minority soil class to improve the models performance. The study area is located in central north Iran with 84 and 72 soil profiles sampled in the donor and recipient areas, respectively. A set of various environmental covariates including remotely sensed data, digital elevation model derivatives and geomorphology map were used as explanatory variables for predicting soil classes. Results showed that among eleven investigated models, C5.0 decision tree (DT), random forest (RF) and multinomial logistic regression (MNL) had the highest overall accuracy of 46%, 42% and 38%, respectively, for the extrapolation of soil classes. Also the Kappa statistic values for these models were 0.30, 0.24 and 0.22, respectively. Oversampling of the minority soil class led to an increase in overall accuracy for some of the models with the highest ones being DT = 53% and RF = 50%. Also, the Kappa value for DT and RF models increased to 0.39 and 0.35, respectively. In addition, oversampling of the minority soil class led to the prevention of losing this class in the final map.