Machine Learning Approaches for Groundwater Salinity Prediction Under Salt Dome Influence in Arid Regions

Groundwater salinization near salt domes is a major threat to water resources in arid and semi-arid plains, causing a persistent decline in water quality worldwide. Machine learning offers an effective approach for evaluating this risk. This study tested three algorithms—Artificial Neural Network (ANN), Support Vector Machine (SVM), and Random Forest (RF)—to estimate groundwater salinity, focusing on the combined roles of land use and climate. The Darab Plain (Iran), where eight salt domes influence groundwater quality, was selected as the study area, and the spatial distribution of the sodium adsorption ratio (SAR) and electrical conductivity (EC) in its groundwater was estimated for 2003, 2013, and 2022, together with their controlling factors. The results showed that the Random Forest model (with an R2 value in 2003 of 0.76 for EC and 0.62 for SAR) achieved the highest accuracy and was selected as the best-performing model. Salinity zoning revealed that the eight salt domes exert a strong direct effect on the salinity of adjacent aquifers and an indirect effect on more distant aquifers along groundwater flow paths. Among the 29 variables examined, dependence-plot analysis of the trained models showed that two collinear terrain-elevation proxies, land surface temperature and channel network base level, most strongly explained both EC and SAR in nearly every modeled year, with temperature and precipitation also consistently influential where climatic data were available. Given the projected warming trend in the region, local authorities and stakeholders should prioritize improved irrigation practices before considering water-management strategies that anticipate reduced agricultural productivity and associated population migration.