Comprehensive Risk Assessment of Sarkhoon Aquifer Salinization Using a Combination of Machine Learning Models

Awareness of salinity risk is of great importance for water resources management, especially in regions near the coast. In this study, it was attempted to develop a comprehensive framework for salinity risk assessment by combining statistical and chemical analysis for Sarkhoon aquifer, Hormozgan province. In the first step, the input layers required for the DRASTIC model were prepared to investigate the aquifer vulnerability and then these layers were combined based on the model procedure to obtain the vulnerability map. Then, the map of salinization hazard occurrence probability was obtained by using three machine learning models of Random Forest, XGBOOST, and BART with considering 12 factors affecting groundwater including topographic wetness, soil, vegetation and other factors. Evaluation of the modeling performance with the receiver operating characteristic (ROC) curve indicated that all three algorithms had very good accuracies with area under curve (AUC) values higher than 90%. Thus, all three models were combined based on their AUC values and the united map for the probability of salinization hazard occurrence was obtained. Finally, the map of salinization risk was generated based on the values for the vulnerability, salinity and hazard occurrence probability. The survey of the risk map showed that the eastern part of the aquifer has high salinization risk due to the high concentration of agricultural land in the area. The results of this study revealed that the achievement of a reliable map for assessing aquifer salinization risk is possible by combining machine learning models.