Investigating the effectiveness of resampling algorithms in improving the classification of unbalanced data in digital soil mapping

In recent years, the use of digital soil mapping (DSM) based on machine learning algorithms with the aim of preparing soil maps has become widespread with the basis of soil class prediction with the help of modeling the relationships between them and environmental variables. One of this method's challenges is the imbalanced nature of soil distribution in landscape, which leads to overfitting and underfitting of classes, and as a result, reduces the accuracy of many used models. This study was conducted to evaluate the ability of two machine learning algorithms, including random forests and support vector machines, for the digital mapping of soil classes with an imbalanced data set. This study was conducted on 95 soil profile classes at the family level, in 4000 hectares of land in the Honam sub-basin, Lorestan province. The issue of imbalance in soil classes was investigated by using six data sets, including the original soil data set and five data sets created by several resampling approaches including two manual classifications and three over-sampling, under-sampling, and Synthetic Minority Over-Sampling Techniques in the R software. The results showed that despite the low values of overall accuracy, the Geographical distribution of soils with high frequency in the study area in digital soil map obtained from the random forest and the original data set as well as Synthetic Minority Over-Sampling Technique, with conventional soil map of study area is significant. Therefore, the low observation number of other soil classes and as a result incorrect training of models can be considered as one of the main reasons for the low accuracy of the used models.