A comparative study of machine learning and Fuzzy-AHP technique to groundwater potential mapping in the data-scarce region

Abstract

Groundwater prospecting with reasonable accuracy is often a challenging task. The integration of geographic information system (GIS) with Machine Learning techniques has proven very reliable to delineate the nonlinear behaviour of groundwater occurrence. The weighted overlay analysis is one of the many essential tools used for groundwater potential zonation, based on remote sensing technology and has been extensively used in the several research work. However, the performance of these methods have not been evaluated for the data-scarce regions. The fuzzy-analytical hierarchy process (fuzzy-AHP) has been used to assign weights to the most used hydrological conditioning parameters to calculate groundwater potential index. Support vector classifier, k-nearest neighbors and random forest classifier have been used as machine learning (ML) algorithms to predict the groundwater potential. The models have been trained, tested and deployed based on 208 yield data of wells. All the affecting parameters have been developed using remotely sensed data. The results have been compared based on precision, recall, f1-score and Cohen's kappa score. As only 208 well data was available for 16997 Km2 of the study area, it has been observed that ML models failed to delineate the groundwater potential zones compared to fuzzy-AHP based weighted overlay analysis. Data augmentation has been used to generate the well data with distributed independent variables, which eventually increased the number of datasets and improved the performance of ML model significantly.