Investigating the competency of some selected soft computing techniques for modeling of lateritic soil strength based on index properties

This study aims to assess the capability of some soft computing techniques including ANN, M5P and RF to accurately predict the strength of selected lateritic soils in southwestern Nigeria from index properties including specific gravity, linear shrinkage, liquid limit, plasticity index, fine sand content, and fines content. To achieve this goal, the experimental dataset obtained from the laboratory analysis of three hundred soil samples taken from thirty different lateritic deposits within southwestern Nigeria was divided into model and gaging dataset. The model dataset contains two hundred and forty data points, which were divided into 70% for training and 15% each for testing and validation of the proposed models. The gaging dataset contains sixty data points, which were used to validate the proposed models against prominent existing models in the literature. The models performances were evaluated using various statistical estimators. Based on the statistical estimators, the proposed models outperformed the existing models in the literature and provided satisfactory performances, thus, they are validated. The obtained R2 values using the ANN model are 0.9967, 0.9963, 0.9989, and 0.9852 for training, testing, validation, and gaging dataset, respectively; the R2 values obtained for M5P model are 0.6676, 0.5501, 0.636 and 0.6727; and the R2 values for RF model are 0.8346, 0.6380, 0.7564, and 0.7901. This implies that ANN provided the most reliable model for the prediction of the soil strength. Thus, ANN is strongly recommended for prediction of lateritic soil strength.