Temporal evaluation of seawater intrusion vulnerability in Shabestar Plain using GALDIT and AI techniques

Abstract

Groundwater contamination, with seawater intrusion (SWI) being the most widespread form particularly in coastal areas, has become a pressing global environmental challenge. Groundwater serves as a vital freshwater resource, particularly in arid and semi-arid regions, making its efficient management essential. In this study, the GALDIT method—an index-based approach that evaluates the vulnerability of aquifers by scoring six key parameters based on expert judgment (groundwater occurrence (G), aquifer hydraulic conductivity (A), groundwater elevation above sea level (L), distance from the shoreline (D), impact of existing seawater intrusion (I), and aquifer thickness (T))—was employed to assess the vulnerability of the Shabestar aquifer to SWI. The study employs the GALDIT method to map aquifer vulnerability for 2002, 2012, and 2022, enabling a temporal comparison of changes over time. The final GALDIT index map, categorized into low, moderate, and high vulnerability classes, revealed an increase in very high vulnerability areas from 10.9% in 2002 to 17.8% in 2022, alongside a decrease in moderate vulnerability areas from 56.4 to 37.3%, indicating a deteriorating condition of the aquifer. However, the reliance on expert judgment introduces potential subjectivity and bias in the vulnerability assessment. To mitigate these limitations, AI-based models, namely artificial neural networks (ANNs) and random forest (RF), were applied to enhance model performance. The GALDIT parameters served as input for the AI models, while observed electrical conductivity (EC), a key indicator of water salinity, and total dissolved solids (TDS), an indicator of drinking water quality, were used as output variables to estimate condition for the year 2022. Results demonstrated that the ANN model outperformed the RF model during verification, improving estimation accuracy by up to 10% and 9% in terms of the determination coefficient (DC), respectively. To further enhance model interpretability and identify the most influential parameters for EC and TDS estimation, a global, variance-based sensitivity analysis using the Sobol method was conducted. This analysis revealed that factors I and D were the most influential for EC, while factors I and T had the greatest impact on TDS in the study region.