Machine learning-based estimation of specific yield in single-well unconfined aquifers

Abstract

Accurate determination of hydrodynamic parameters holds immense significance in achieving sustainable development objectives and ensuring comprehensive water resources management. This research aims to assess the performance of machine learning models in estimating specific yield using hydrogeological data and pumping test data in the single well system for the unconfined alluvial aquifers of Iran’s central plateau. Artificial neural network, gene expression programming, adaptive neuro-fuzzy inference system, fuzzy logic, least square support vector machine, and group method of data handling were used to estimate specific yield. In order to accomplish this objective, data from pumping tests and hydrogeological evaluations conducted on 83 wells located in the central plateau of Iran were gathered and normalized. The available data was utilized to generate different combinations as input for the models. Subsequently, 70% of the data was selected for the training phase, while the remaining 30% was allocated for the testing phase. The precision of the models was evaluated by employing mean absolute error, root mean square error, and correlation coefficient. The performance criteria indicated that while all of the machine learning models are promising tools for rapid and dependable specific yield estimation, the group method of data handling model exhibited the highest level of accuracy. The results indicated that combination 3 (well discharge, the slope of the initial and final sections of the time-drawdown graph, minimum and maximum drawdown data, water depth, the distance of the observation well from the pumping well, and hydraulic conductivity) was the most effective in estimating specific yield.