K-Fold and State-of-the-Art Metaheuristic Machine Learning Approaches for Groundwater Potential Modelling

Abstract

Groundwater being an essential resource is not easily available in some parts of the world. The present study, aimed at procuring better prediction maps for groundwater potential zones, is based on a novel approach combining the use of k-fold cross-validation method and the implementation of four scenarios, each comprising of six machine learning models, ANFIS (Adaptive Neuro Fuzzy Inference System) and five other ensembles of it, ANFIS-Firefly, ANFIS-Bees, ANFIS-GA, ANFIS-DE and ANFIS-ACO. Ada Boost Model has played a vital role in determining the collinearity among the fourteen conditioning factors, which are, Lithology, Slope, TST, TRI, LULC, HAND, Curvature, Distance to Stream, Distance to Fault, Rainfall, Fault Density, Drainage Density, Elevation and Aspect. The AUCROC (Area Under Curve – Receiver Operating Characteristics) approach was employed as a model evaluation metric along with Accuracy, Sensitivity and Specificity. Among the models, ANFIS-DE showed the most promising results, acquiring the highest average values among the four scenarios for AUC (0.934), Accuracy (0.987), Sensitivity (0.985) and Specificity (0.985). Promising results of this study gives the necessary incentive for further applying this approach for groundwater zonation of other areas of the world as well as other areas of hydrogeological studies.