Application of machine learning techniques for predicting seawater intrusion vulnerability in a coastal alluvial aquifer of eastern India.

Groundwater contamination is a major concern worldwide, especially in coastal regions, due to increasing seawater intrusion into aquifers. In eastern India, increasing salinization of coastal aquifers necessitates robust predictive tools for assessing seawater intrusion vulnerability and supporting sustainable groundwater management. This study evaluates the performance of three machine-learning (ML) techniques viz., Random Forest (RF), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM), in predicting seasonal seawater intrusion vulnerability in a coastal river basin of eastern India. Seasonal groundwater-level and hydrochemical data [Electrical Conductivity (EC), Cl-, and HCO3-] of leaky confined aquifer, spanning 2012-2021, together with the pumping-test and lithological data, were utilized for the model development. Five hydrogeological predictors such as 'Aquifer Hydraulic Conductivity (A)', 'Groundwater Elevation (L)', 'Distance from the Coastline (D)', 'Extent of Seawater Intrusion (I)', and 'Aquifer Thickness (T)' were employed as "input" variables, whereas 'adjusted SWIVI' values derived from a modified GALDIT framework served as the "target" variable. The results revealed that the modified GALDIT method successfully delineated seawater intrusion-prone zones, exhibiting moderate correlations with EC (r = 0.552-0.577) and good AUC values (0.862-0.865). Among the evaluated ML models, the LSTM technique demonstrated the highest predictive accuracy, yielding the lowest MAE (0.0141-0.0170) and RMSE (0.0184-0.0216) and the highest correlations (r = 0.9805-0.9910) and model efficiencies (NSE = 0.9524-0.9632) during both training and testing phases. Although the RF and SVM models yielded slightly low PBIAS, the LSTM technique consistently yielded superior overall performance and the closest spatio-temporal agreement with 'adjusted SWIVI'. These results highlight the ability of ML models to capture complex hydrogeological interactions governing seawater intrusion and demonstrate the potential of LSTM as a reliable tool for assessing seawater intrusion vulnerability and proactive groundwater management under data-scarce coastal aquifer systems.