Spatial prediction of WRB soil classes in an arid floodplain using multinomial logistic regression and random forest models, south-east of Iran

Abstract

In the current study, the variations of soil classes at the first and second levels of WRB (World Reference Base for soil resource) soil classification system were investigated by two machine learning including multinomial logistic regression (MLR) and random forest (RF) models in an arid floodplain which covers an area approximately 600 km2 located in Sistan region, Iran. The model’s performance was tested using 10-fold cross-validation by calculation of overall model accuracy and the kappa statistic. Three main Reference Soil Groups (RSGs) including Cambisols, Fluvisols, and Solonchaks at the first level, and 18 WRB soil groups at the second level were identified. Results showed that the overall accuracy at the first level of WRB was 53% and 49% with a kappa of 0.26 and 0.19 for MLR and RF models, respectively. At the second level of WRB, the overall accuracy was 11% and 21% with a kappa of 0 and 0.09 for MLR and RF models, respectively. Also, results showed that the MLR model had better performance (overall accuracy = 53%) at the first level of WRB, but the RF model showed better prediction (overall accuracy = 21%) at the second level of WRB. Multiresolution Valley Bottom Flatness Index (MrVBF), Normalized Difference Salinity Index (NDSI), Multiresolution of Ridge Top Flatness Index (MrRTF), convergence index, and channel network base level were among top covariates used for prediction at two levels of WRB. Results revealed the complexity of soil variations in this floodplain. Using other covariates such as soil texture and salinity maps can improve the prediction power. Increasing the size of sampling is recommended to improve the accuracy of the models in predicting the second level of WRB in this area.