Enhancing the accuracy of seawater intrusion vulnerability assessment using a hybrid GALDIT framework in tropical low-lying coastal settings.

Low-lying coastal plains are highly susceptible to seawater intrusion (SWI), driven by factors such as sea-level rise, coastal flooding, and saline encroachment. These challenges are further intensified by anthropogenic activities, which traditional GALDIT models are unable to represent adequately. The GALDIT framework was refined by incorporating 'soil media' and 'well density' layers. Parameter weights and rankings were optimized using the Analytic Hierarchy Process (AHP) and the Wilcoxon signed-rank test, respectively. Despite their utility, both GALDIT and modified (Mod)-GALDIT models are limited in capturing the nonlinear spatial complexities of SWI. To address these limitations, this study introduces a hybrid approach (CNN-XGBoost) for assessing the Seawater Vulnerability Index (SVI) in coastal regions of Odisha. The method integrates Mod-GALDIT with hybrid deep learning techniques, specifically Convolutional Neural Networks (CNN) for feature extraction and Extreme Gradient Boosting (XGBoost) for predictive modelling. Model validation using Receiver Operating Characteristic (ROC) curves and Total Dissolved Solids (TDS) data, resulting in Area Under the Curve (AUC) values of 85%, 81%, 78%, and 74% for the CNN-XGBoost-GALDIT, CNN-GALDIT, Mod-GALDIT, and traditional GALDIT models, respectively. The CNN-XGBoost model (R2 = 94%) outperformed the standalone CNN model (R2 = 90%) by mitigating overfitting through XGBoost's regularization techniques. Spatial analysis revealed the highest vulnerability in the eastern part (47.3%) and lower vulnerability in the western inland regions (45%), demonstrating clear spatial improvement. This study pioneers the large-scale application of hybrid models for SVI assessment, enhancing accuracy and offering robust solutions to global challenges of groundwater depletion, salinity intrusion, and coastal water management.