Abstract
The present study aimed to create novel hybrid models to produce groundwater potentiality models (GWP) in the Teesta River basin of Bangladesh. Six ensemble machine learning (EML) algorithms, such as random forest (RF), random subspace, dagging, bagging, naive Bayes tree (NBT), and stacking, coupled with fuzzy logic (FL) models and a ROC-based weighting approach have been used for creating hybrid models integrated GWP. The GWP was then verified using both parametric and nonparametric receiver operating characteristic curves (ROC), such as the empirical ROC (eROC) and the binormal ROC curve (bROC). We conducted an RF-based sensitivity analysis to compute the relevancy of the conditioning variables for GWP modeling. The very high and high groundwater potential regions were predicted as 831-1200 km(2) and 521-680 km(2) areas based on six EML models. Based on the area under the curve of the ROC, the NBT (eROC: 0.892; bROC: 0.928) model outperforms rest of the models. Six GPMs were considered variables for the next step and turned into crisp fuzzy layers using the fuzzy membership function, and the ROC-based weighting approach. Subsequently four fuzzy logic operators were used to assimilate the crisp fuzzy layers, including AND, OR, GAMMA0.8, and GAMMA 0.9, as well as GAMMA0.9. Thus, we created four hybrid models using FL model. The results of the eROC and bROC curve showed that GAMMA 0.9 operator outperformed other fuzzy operators-based GPMs in terms of accuracy. According to the validation outcomes, four hybrid models outperformed six EML models in terms of performance. The present study will aid in enhancing the efficiency of GPMs in preparing viable planning for groundwater management.