Advancing Groundwater Vulnerability Assessment in Coastal Regions: Integrating Machine Learning and Traditional Frameworks

We introduce an innovative machine learning (ML)-enhanced method to assess groundwater vulnerability in coastal regions, with a specific focus on the Azarshahr plain near Urmia Lake in Northwestern Iran. Our methodology integrates the traditional DRASTIC and GALDIT frameworks to surpass their limitations (e.g. subjectivity) in varied contexts such as coastal and agricultural-industrial environments. The traditional frameworks including the DRASTIC framework form the core of our approach, featuring seven key layers: Depth to water [D], Net Recharge [R], Aquifer Media [A], Soil Media [S], Topography [T], Impact of Vadose Zone [I], and Hydraulic Conductivity [C], each meticulously developed with specific ratings and weights according to DRASTIC standards. Similarly, the GALDIT framework contributes a six-layer map, including Groundwater Occurrence [G], Aquifer Hydraulic Conductivity [A], Height of Groundwater Level [L], Distance from the Shore [D], Impact of Existing Seawater Intrusion Status [I], and Aquifer Thickness [T], each layer uniquely rated and weighted. To address the limitations of these traditional frameworks, our study integrates an advanced ML recalibration of the GALDIT and DRASTIC indices, using the maximum concentrations of Total Dissolved Solids (TDS) and Nitrate (NO₃) in the study area as proxies. We employed a range of decision tree-based ML models, including Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGBM), and Random Forest (RF), to predict the adjusted vulnerability indices, applying six predictors for GALDIT and seven for DRASTIC. These models were trained and validated on a dataset split into 70% for training and 30% for validation. Our results indicate that the traditional DRASTIC indices correlate weakly with NO₃ concentrations. However, the ML-augmented models, particularly AdaBoost, significantly improved predictive accuracy. Likewise, GALDIT results were greatly enhanced by incorporating the AdaBoost model. A key innovation in our research is the development of a sophisticated meta-ensemble ML model. This model, based on the most effective AdaBoost applications in the DRASTIC and GALDIT assessments, marks a significant methodological advancement. It integrates vulnerabilities from both frameworks using a Fuzzy operation and then redeveloping a meta-ensemble ML model. This comprehensive model demonstrated exceptional performance, highlighting the effectiveness of our integrated ML approach in providing a more detailed, accurate, and robust assessment of coastal aquifer vulnerability. Moreover, our study includes an extensive spatial analysis of groundwater vulnerability in the Azarshahr plain. The DRASTIC model indicated varying vulnerability levels, with heightened susceptibility in central and southern regions, albeit showing a weaker correlation with NO₃ concentrations. Conversely, AdaBoost exhibited a strong correlation with actual NO₃ levels, showcasing its predictive capability. The GALDIT index identified several high-risk areas, particularly those vulnerable to seawater intrusion, with the AdaBoost-enhanced model outperforming other ML approaches. Our comprehensive AdaBoost meta-ensemble model merges insights from both NO₃ and TDS evaluations, offering a holistic groundwater vulnerability. This model is crucial for informed decision-making, identifying areas where NO₃ and TDS risks converge. Its spatial analysis strongly correlates 'Very High' vulnerability zones with high NO₃ and TDS concentrations, confirming its integrative efficiency in environmental risk assessment.