An improved GALDIT method combined with machine learning for assessing aquifer vulnerability to seawater intrusion in the Shandong Peninsula, China

The GALDIT method is one of the most prevalent methodologies for assessing seawater intrusion vulnerability. However, the subjectivity of the vulnerability assessment framework and the complexity of the factors influencing seawater intrusion pose challenges to accurate mapping of vulnerability assessment. Hence, this paper proposes a new vulnerability assessment model for seawater intrusion based on the GALDIT method, incorporating machine learning techniques (Artificial Neural Networks, ANN, and Random Forests, RF) and triangular fuzzy membership functions (FMF). The new modelling framework introduces “Water yield property of the aquifer” for representing the influence of geological structures on groundwater storage status and adds a "Land Use type" factor to characterize the impact of human activities, and is referred to as "WALDIT_LU". This framework was tested in a coastal aquifer in Shandong Province, China. The results show that the thematic maps improved by the FMF method are more objective and better suited for regions with extensive data ranges or scales than those produced by the original GALDIT method. Hydrochemical validation results indicate a significant enhancement in the accuracy of vulnerability maps created by the WALDIT_LU-ANN and WALDIT_LU-RF models compared to the original GALDIT model. The Spearman’s rank correlation coefficient values obtained between the GALDIT, WALDIT_LU-ANN, WALDIT_LU-RF and the Cl- ion were 0.291, 0.426 and 0.477, respectively. The equivalent ratio values using the TDS as the parameter were 0.275, 0.737 and 0.811, respectively. The optimised factor weights for the WALDIT_LU-RF model are more reasonable with factor weights of 25.52% (I), 14.47% (A), 14.38% (D), 12.49% (T), 11.73% (LU), 10.68% (W), and 10.73% (L). It is concluded that the new framework incorporating the WALDIT_LU index provides a more comprehensive consideration of the factors influencing seawater intrusion. Additionally, the new model reduces subjectivity and enhances the reliability of mapping seawater intrusion vulnerability.