Introduction
Groundwater salinity represents one of the most serious threats to water quality in arid and semi-arid regions, directly influencing agriculture, ecosystems, and sustainable development. In such vulnerable environments, identifying and mapping areas susceptible to salinization are essential for effective water resource management and informed land-use planning. Salinization reduces the availability of potable water and degrades soil quality, leading to declining crop yields and long-term ecological imbalance. Increasing dependence on groundwater caused by population growth, agricultural intensification, and industrial development further aggravates the problem, especially in regions with low recharge rates and high evaporation. Iran, located mainly within arid and semi-arid climatic zones, faces considerable challenges in maintaining the quality and sustainability of its groundwater resources. The Koohpayeh-Segzi plain in Isfahan Province exemplifies these challenges, as groundwater plays a fundamental role in sustaining agricultural productivity and local livelihoods. However, extensive groundwater extraction combined with natural geochemical processes has resulted in a steady increase in salinity levels. Consequently, understanding the spatial distribution and controlling factors of groundwater salinity in this region is crucial for mitigating future risks. This study, therefore, aims to evaluate groundwater salinity susceptibility in the Koohpayeh-Segzi plain using advanced machine learning techniques to improve predictive accuracy and support sustainable groundwater management strategies.
Materials and Methods
This study employed two ensemble learning algorithms Adaptive Boosting (AdaBoost) and Bagged AdaBoost to evaluate groundwater salinity susceptibility in the Koohpayeh-Segzi plain. The Bagged AdaBoost model represents an enhanced version of the standard AdaBoost algorithm, incorporating bootstrap-based aggregation to improve model robustness and predictive reliability. The dataset used for modeling consisted of annual average salinity observations from 50 monitoring wells recorded over a 23-year period, providing a comprehensive temporal representation of groundwater quality dynamics. A wide range of conditioning factors was considered as predictor variables, encompassing topographic parameters (elevation, slope, and aspect), climatic variables (evaporation and precipitation), hydrological indices (topographic wetness index and distance to streams), hydrogeological indicators (depth to groundwater table and groundwater level decline), geological factors (distance to faults and lithology), as well as soil order and land use types. All spatial data layers were prepared and standardized in a geographic information system (GIS) environment to ensure consistency across scales and units. Model performance was quantitatively assessed using multiple statistical metrics, including accuracy, precision, Kappa coefficient, and F1-score, to ensure reliable evaluation of classification outcomes. The final groundwater salinity susceptibility maps were produced based on the trained ensemble models, illustrating the spatial distribution of salinity risk across the study area and offering critical insights for sustainable groundwater management and regional land-use planning.
Results and Discussion
The comparative analysis of model performance demonstrated that the Bagged AdaBoost algorithm significantly outperformed the standard AdaBoost across all evaluation metrics, indicating its superior capability in capturing complex patterns associated with groundwater salinity. Specifically, the overall accuracy increased from 0.89 to 0.93, precision improved from 0.67 to 0.80, F1-score rose from 0.80 to 0.89, and the Kappa coefficient a measure of agreement beyond chance enhanced from 0.72 to 0.85. These improvements reflect the enhanced stability and generalization power of the Bagged AdaBoost model, particularly in handling heterogeneous environmental data. To further interpret model behavior, a variable importance analysis was conducted, revealing that groundwater depth, elevation, and evaporation were the most influential predictors in determining salinity susceptibility. These variables are closely linked to the region’s hydrogeological and climatic conditions, underscoring their critical role in salinization processes. The spatial susceptibility map generated from the optimized model illustrated a distinct gradient in salinity risk, with elevated levels predominantly concentrated in the southern and western portions of the Koohpayeh-Segzi plain. In contrast, the northern and eastern zones exhibited relatively lower susceptibility. This spatial pattern corresponds well with known regional dynamics, including groundwater flow direction, recharge limitations, and anthropogenic pressures such as intensive agricultural activity and land-use changes. The findings highlight the utility of ensemble learning approaches in environmental modeling and provide actionable insights for targeted groundwater management and salinity mitigation strategies in vulnerable arid and semi-arid regions.
Conclusion
The integration of Adaptive Boosting (AdaBoost) with bagging techniques substantially enhances the robustness, accuracy, and predictive reliability of groundwater salinity susceptibility modeling, particularly in regions characterized by data scarcity and environmental heterogeneity. By combining AdaBoost’s iterative error-correction capability with bagging’s variance-reduction mechanism, the hybrid Bagged AdaBoost model achieves greater stability, minimizes overfitting, and demonstrates improved generalization across diverse datasets. The generated groundwater salinity susceptibility maps provide detailed spatial insights into areas most prone to salinization, offering valuable information for water resource managers, agricultural planners, and environmental policymakers. These maps enable the identification and prioritization of critical zones requiring immediate intervention, thus supporting the design of adaptive and site-specific management strategies aimed at mitigating salinity risks. Moreover, the results highlight the effectiveness of ensemble-based machine learning approaches in capturing complex nonlinear relationships among environmental, geological, and hydrological factors. The study also emphasizes the importance of integrating machine learning frameworks with geographic information systems to enhance visualization, interpretation, and practical applicability of model outputs. Overall, this research demonstrates the strong potential of ensemble learning models for groundwater quality assessment and advocates for their broader application in arid and semi-arid regions, where conventional statistical or deterministic methods often face limitations due to insufficient, inconsistent, or highly variable datasets. |
