Assessment of groundwater potential modeling using support vector machine optimization based on Bayesian multi-objective hyperparameter algorithm

Today, water supply in order to achieve sustainable development goals is one of the most important concerns and challenges in most countries. For this reason, accurate identification of areas with groundwater potential is one of the important tools in the protection, management and exploitation of water resources. Accordingly, the present study was conducted with the aim of modeling and predicting groundwater potential in Markazi province, Iran using Multivariate adaptive regression spline (MARS) and Support vector machine (SVM) machine learning models and using two random search (RS) and Bayesian optimization hyperparameter algorithms to optimize the parameters of the SVM model. For this purpose, 18 variables affecting the groundwater potential and 3482 spring locations were used to model the groundwater potential. Data for modeling were divided into two categories of training (70%) and validation (30%). The receiver operating characteristics (ROC) were used to evaluate the performance of the models. The results of evaluation models showed that using hyperparameters random search and Bayesian optimization were improved SVM accuracy in training and validation stages. Bayesian optimization methods are very efficient because they are consciously choosing the parameters of the model that this strategy improves the performance of the model. Evaluating accuracy in the validation stage showed that the AUC value is for MARS, SVM, RS-SVM and B-SVM models 87.40%, 88.25%, 90.73% and 91.73%, respectively. The results of assessment variables importance showed elevation, precipitation in the coldest month, soil and slope variables have the most importance in modeling groundwater potential, while aspect, profile curvature and TWI variables, have the least importance in predicting groundwater potential in Markazi province.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Water supply is one of the most important concerns and challenges for achieving sustainable development goals in most countries. Accurate identification of areas with groundwater potential is therefore crucial for the protection, management, and exploitation of water resources. This study used Multivariate adaptive regression spline (MARS) and Support vector machine (SVM) machine learning models, along with two random search (RS) and Bayesian optimization hyperparameter algorithms, to model and predict groundwater potential in Markazi province, Iran. The results showed that using hyperparameters random search and Bayesian optimization improved the accuracy of the SVM model in both training and validation stages. The evaluation of accuracy in the validation stage revealed AUC values of 87.40%, 88.25%, 90.73%, and 91.73% for the MARS, SVM, RS-SVM, and B-SVM models, respectively. The assessment of variable importance indicated that elevation, precipitation in the coldest month, soil, and slope variables were the most important in modeling groundwater potential, while aspect, profile curvature, and TWI variables had the least importance in predicting groundwater potential in Markazi province.