Modeling Groundwater Potential Using Machine Learning Models

Finding the potential of groundwater resources is one of the basic principles in water resources management. The aim of this research is to determine the potential of groundwater using support vector machine learning (SVM) models as well as metaheuristic algorithms (hybrid support vector machine model and the bee metaheuristic optimization algorithm (SVM-BA) and hybrid model of the support vector machine and particle swarm optimization algorithm (SVM-PSO).

The factors of elevation, slope, aspect, topographic humidity index, distance from stream, drainage density, distance from fault, lithology, topographic position index, land roughness index, relative slope position and flow convergence index were selected in Bojnurd region. Information on the location of 359 springs was received from the regional water company. Random division algorithm was used to divide training points (70%) and validation points (30%). Based on the removal sensitivity analysis, the importance and contribution of the input variables in determining the groundwater potential were determined. The accuracy of the models was evaluated in two stages of training and validation based on the receiver operating characteristic (ROC) curve method.

The evaluation of the accuracy of the models based on the evaluation criteria of the area under curve (AUC) showed that the prediction accuracy of the hybrid model of the support vector machine and the particle swarm optimization algorithm (SVM-PSO) is 0.945 more than other models (SVM: 0.918 and SVM-BA: 0.932). Based on the results of the superior model, the high potential class and the very high potential class accounted for 7.75% and 38.66% of the area respectively. Among the factors, relative slope position with 14.5%, distance from the fault with 13.4% and lithology with 12.3% were the most important in predicting groundwater potential.

Based on the results of this research, the support vector machine model has a high performance, and two optimization algorithms, the bee metaheuristic and particle swarm optimization algorithm, strengthen the predictive power of the model. Also machine learning models can identify the relationship between the environmental factors and the water supply of the springs and determine their role by using the available data. The relative slope position factor was identified as the most important variable and the distance from the fault factor was considered as the second most important variable in the present study. The results of the research showed that the faults in the region play an important role in aquifer recharge, storage and flow of groundwater. The lithological factor was also introduced by the model as the third important variable in identifying the state of groundwater potential. In this research, by presenting the groundwater potential map, it is possible to plan and verify land use planning for the Bojnurd watershed.