Fig. 2. A flowchart of MLP-GWO. Fig. 3. Location of the sampling map in the southwest of Iran (URL 1). Correlation matrix of the data used. Table 2 Statistical characteristics of data in the Khuzestan Province, southwest of Iran. Table 1 Fig. 4. The variations of different soil parameters and EC values. The performance evaluation criteria of developed models. The optimal MLP-GWO hybrid and the comparative standalone MLP model are indicated in the boldfacec row Fig. 5. Schematic structure of the hybrid MLP-GWO model developed for EC prediction. Table 3 Fig. 6. The observed and predicted electrical conductivity (EC) in the testing phase by optimal (i.e., MLP-GWO and MLP) models. Fig. 7. Scatter plots of measured and estimated values of EC for the optimal models (MLP4 and MLP-GWO4). Fig. 8. Residual error of estimated electrical conductivity (EC) for the best models (i.e., MLP4 and MLP-GWO4) in the testing phase In predictive modelling, a substantial challenge confronted by ar- tificial intelligence methods is searching optimal weights in a neuronal layer to facilitate the extraction of appropriate input features in the input data as requirements for creating an ideal predictive model. For It is interesting to note that in the second set of models (i.e., MLP- GWO2 & MLP2), the elimination of the OM from the input variable list ed to a mild increase in the error of EC prediction. In that case, al- hough the prediction accuracy was lower the equivalent model in the irst set (i.e., MLP-GWO1), the RMSE and RRMSE values of the hybrid MILP-GWO2 were less than corresponding values of the standalone MILP2 model. In the other words, using the GWO as an optimizer al- zorithm has led to improved performance in terms of the RMSE and RRMSE values of the standalone MLP model by 20.0% and 16.3%, re- spectively. A closer examination of these results shows that the input variables in the third set of models (i.e., MLP-GWO3 & MLP3) resulted n the lowest prediction accuracy among all of the other input combi- 1ations. It is therefore conclusive that the hybrid MLP-GWO3 model hat included the input variables denoted as Ca*”, Mg*?, Na*, Cl”, $0, 7, HCO3” was unable to attain a low prediction error. Moreover, it Fig. 9. Histograms of different electrical conductivity (EC) values during the testing period.
