Comparative Evaluation of Decision Tree (M5) and Least Square Support Vector Machine (LS-SVM) Models for Groundwater Level Prediction in the Mashhad Plain

Objective: Predicting the groundwater table level is considered one of the fundamental steps in the optimal management of water resources in arid and semi-arid regions. Nowadays, the application of intelligent models for estimating groundwater levels is increasing due to their ease of use and high accuracy in estimating complex and nonlinear mathematical equations. The aim of the present study is to estimate the groundwater table level of the Mashhad plain aquifer using the decision tree model (M5) and to compare it with the least squares support vector machine model (LS-SVM) under 10 different scenarios.Method: For this purpose, monthly climatic data (precipitation, evaporation, and temperature) and groundwater level information from 60 piezometric wells over a 10-year statistical period were utilized, and the employed models were evaluated using statistics such as the coefficient of determination (R2), RMSE, and MBE.Results: The results of the LS-SVM model indicated that the highest simulation accuracy belonged to scenario 4, followed by scenario 9, while the other scenarios exhibited very low accuracy in simulating the water level. The MBE error values in scenarios 4 (-0/151) and 9 (-0/018) showed that the model simulated the groundwater level lower than reality. Based on the results of the water level simulation using the decision tree model, all scenarios were acceptable, and the scenarios 4 and 5 having the highest and lowest accuracy with coefficient of determination of 0/999 and 0/86, respectively. Overall, in both models used, scenario 4 simulated the groundwater level with almost similar accuracy. A comparison of the results of the models indicated that the LS-SVM model is more sensitive to changes in input parameters than the M5 model, such that the decision tree model, unlike the least squares support vector machine model, provided acceptable results in all scenarios.Conclusions: In summary, the comparison of the models used suggests that the appropriate selection of climatic parameters and the examination and analysis of data have a significant impact on the accuracy of predictions.