A novel data-driven framework combining graph theory and machine learning to predict groundwater quality: Evidence from the Lake Urmia aquifer

Groundwater quality monitoring and sustainable management are crucial for long-term water security, especially in arid and semi-arid regions impacted by climate change and human activities. This study investigates the relationship between spatial graph-theoretical metrics and groundwater quality parameters in the Urmia aquifer (Iran) using advanced machine learning algorithms, including Random Forest (RF), Support Vector Regression (SVR), and Linear Regression (LR). The results indicate that the graph-theoretical metrics are more influential in this aquifer. Results showed that RF achieved R 2 values greater than 0.90 for all parameters (superior to SVR, which had R 2 > 0.70 for most, and LR, which had R 2 < 0.50), with connectivity metrics, especially SPLIT, dBC_IIC, and PD, accounting for over 85% of the variance in salinity indicators. At the same time, for the cations (K + , Na + , Mg 2+ , and Ca 2+ ) during both the training and testing periods, the R 2 values exceeded 60%. RF outperformed SVR in predicting the HCO 3 − and CO 3 2− anions. This framework provides a transferable tool for spatially explicit groundwater quality prediction, enabling the early identification of vulnerable zones, targeted land-use planning to mitigate salinity, and informed conservation strategies for policymakers in similar arid aquifers worldwide, thereby supporting sustainable water resource management and global water security.