Use of Gamma test technique for choosing the optimum input variables in modeling of soil shear strength using artificial neural networks

Gamma test is an appropriate tool for determining the optimum input combination and suitable number of data to achieve minimum mean square error in any continuous nonlinear modeling approaches. In this study, at first Gamma test technique was used to determine the optimum input variables from measured input parameters (including soil properties, topographic and vegetation attributes) affecting soil shear strength (SSS) prediction. Two different artificial neural network (ANN) models were then constructed using different input data (Models 1 and 2) to predict SSS. In Model 1, all of the measured parameters (12 parameters) were used as input variables of the model and Model 2 was constructed using only the optimum 5 parameters resulted from the Gamma test trails. According to the Gamma test trail results, fine sand and mean weight diameter (MWD) parameters had the lowest and highest Gamma and v-ratio values among the other parameters, respectively. Furthermore, the vegetation index (NDVI), sand, very fine sand, and aspect parameters had a Gamma values of 0.2177, 0.2280, 0.2313, and 0.2318 in comparison with the other investigated parameters, respectively. Therefore, sand, fine sand, very fine sand, NDVI, and aspect parameters were selected as the optimum 5 input variables for modeling of SSS using ANNs. The proposed ANN model using the optimum 5 input variables, selected from the Gamma test trails (Model 2), had a similar accuracy with the proposed ANN model using all of the 12 input variables (Model 1). The correlation coefficient (r) and root mean square error (RMSE) values for Model 2 were 0.885 and 0.045, respectively, while these indices for Model 1 were 0.891 and 0.058, respectively. Therefore, it appears that the Gamma test technique can be used for choosing optimum input variables affecting SSS prediction to reduce experimental expenses and to save a great amount of time and effort in modeling approaches.