Comparison of data mining techniques to predict and map the atterberg limits in central plateau of Iran

Abstract

The Atterberg limits display soil mechanical behavior and, therefore, can be so important for topics related to soil management. The aim of the research was to investigate the spatial variability of the Atterberg limits using three most common digital soil-mapping techniques, the pool of easy-to-obtain environmental variables and 85 soil samples in central Iran. The results showed that the maximum amount of liquid limit (LL) and plastic limit (PL) were obtained in the central, eastern and southeastern parts of the study area where the soil textural classes were loam and clay loam. The minimum amount of LL and PL were related to the northwestern parts of the study area, adjacent to the mountain regions, where the samples had high levels of sand content (>80%). The ranges of plasticity index (PI) in the study area were obtained between 0.01 to 4%. According to the leave-in-out cross-validation method, it should be highlighted the combination of artificial bee colony algorithm (ABC) and artificial neural network (ANN) techniques were the best model to predict the Atterberg limits in the study area, compared to the support vector machine and regression tree model. For instance, ABC-ANN could predict PI with RMSE, R2 and ME of 0.23, 0.91 and-0.03, respectively. Our finding generally indicated that the proposed method can explain the most of variations of the Atterberg limits in the study area, and it could be recommended, therefore, as an indirect approach to assess soil mechanical properties in the arid regions, where the soil survey/sampling is difficult to undertake.