Optimization of Statistical and Machines Learning Hybrid Models for Groundwater Potential Mapping

Abstract

Determining areas of high groundwater potential is important for exploitation, management, and protection of water resources. This study assesses the spatial distribution of groundwater potential in the Zarrineh Rood watershed of Kurdistan Province, Iran using combinations of five statistical and machine learning algorithms – frequency ratio (FR), radial basis function (RBF), index of entropy (IOE), evidential belief function (EBF) and fuzzy art map (FAM). To accomplish this, 1448 well locations in the study area were randomly divided into two data sets for training (70%= 1013 locations) and validation (30%= 435 locations) based on the holdout method. Fourteen factors that can affect the presence or absence of groundwater were identified, measured, and mapped using ArcGIS and SAGA-GIS software. The models were used to predict the locations of groundwater based on suitable combinations of the conditioning factors to produce groundwater potential maps. The probability of groundwater at any location was classified as low, moderate, high, or very high based on natural breaks in the data spectrum. The model predictions were tested for validity and their success was determined using receiver operating characteristic (ROC) curves, standard errors (SE), positive predictive value (PPV), negative predictive value (NPV), sensitivity (SST), specificity (SPF) and accuracy (ACC), and the Friedman test. The performance assessments of groundwater potential predictions using the area under the curve (AUC) and accuracy (ACC) showed that the FR-RBF model had very good performance (AUC= 0.889, ACC= 87.51). And FR-FAM (AUC= 0.869, ACC= 84.67), EBF-FAM (AUC= 0.864, ACC= 84.42), EBF-RBF (AUC= 0.854, ACC= 83.94), FR-IOE (AUC= 0.836, ACC= 83.62), and EBF-IOE (AUC= 0.833, ACC= 80.42) also had acceptable performance. The results of the Friedman test also show that there are significant differences between the models and the highest mean rank was generated by the FR-FAM model (3.642). Therefore, the hybrid models can be used to increase the accuracy of groundwater-prediction models in the study region and perhaps in similar settings.