The use of hybrid machine learning models for improving the GALDIT model for coastal aquifer vulnerability mapping

The objective of this study was to improve the predictability of the GALDIT (G: groundwater occurrence, A: aquifer hydraulic conductivity, L: level of groundwater above sea level, D: distance from the shoreline, I: impact of the seawater intrusion, and T: thickness of the aquifer) groundwater vulnerability model using machine leaning methods. This study evaluated eight state-of-the-art machine learning methods, including the naive Bayes tree (NBT) and logistic model tree (LMT) methods, and their combinations with the dagging (DA), bagging (BA), and random subspace (RS) methods. The results of the machine leaning methods were compared against the benchmark GALDIT model. The coastal Gharesoo-Gorgan Rood aquifer, North Iran, was used as a case study for the proposed methodology. Two sets of total dissolved solids (TDS) samples from 53 wells were collected in 2017 and 2018 and used for the GALDIT modeling and validation purposes, respectively. Correlation coefficient (r) values were calculated for model validation and prediction accuracy by comparison with the TDS data. All eight machine learning models performed well in assessing the coastal aquifer vulnerability with respect to the GALDIT model. The best result was obtained by the BA-LMT model (r = 0.931), followed by the DA-LMT model (r = 0.911), the BA-NBT model (r = 0.904), the DA-NBT model (r = 0.896), the RS-NBT model (r = 0.882), the RS-LMT (r = 0.873), the LMT (r = 0.863), the NBT (r = 0.850), and GALDIT model (r = 0.480).