Insight into groundwater level prediction with feature effectiveness: comparison of machine learning and numerical methods

ABSTRACT Reliable simulation of groundwater fluctuation is imperative for implementing sustainable groundwater management so that not only local people will be covered, but also environmental threats are going to be restrained in the future. This research aims to scrutinize the accuracy of numerical model (NM) and machine learning (ML) methods for groundwater level prediction (GWLP) for Yazd-Ardakan Plain from 2012 to 2019 in monthly steps. Additionally, principal component analysis (PCA) is applied to investigate the impact of each ML input feature on GWLP. The study area's aquifer data were analyzed and prepared to develop the conceptual model of MODFLOW and train ML algorithms for GWLP. Considering observation wells (OBWs), operation wells (OPWs), and their latitude and longitude as input features in convolutional neural networks (CNN), support vector machine (SVM), and decision tree (DT) algorithms, GWLP was performed. The results demonstrate that although MODFLOW considers the unique features of the aquifer, the most accurate GWLP was achieved by SVM, with root mean square errors (RMSE), correlation coefficient (), and area under the receiver operating characteristics (ROC) curve (AUC) values of 0.12, 0.90, and 0.94, respectively. Furthermore, PCA presented that the observed groundwater level (GWL) was the most effective feature with 71%.