Document Type : Research Paper
Groundwater vulnerability assessment is an important tool for protecting aquifers against contamination, particularly in coastal and agricultural regions where groundwater quality is affected by both natural and human-induced factors. The DRASTIC index is one of the most widely used methods for evaluating intrinsic groundwater vulnerability. It integrates seven hydrogeological parameters, including depth to groundwater, net recharge, aquifer media, soil media, topography, impact of the vadose zone, and hydraulic conductivity. However, the classical DRASTIC model usually applies fixed weights and ratings, while some of its parameters are temporally variable. Among them, depth to groundwater and net recharge are strongly affected by rainfall variability, groundwater abstraction, and hydrological conditions. Ignoring this temporal variability may reduce the reliability of vulnerability maps, especially in shallow coastal aquifers.
In this study, the effect of temporal uncertainty on the calibration of the DRASTIC index was investigated in the Sari–Neka aquifer. Nitrate concentration was used as the groundwater quality indicator for calibration. The study aimed to evaluate vulnerability changes under different temporal conditions and to compare the performance of two data-driven calibration models, namely RF-HHO and CNN-HHO.
In this study, the standard DRASTIC index was first calculated using seven hydrogeological parameters in a GIS environment. Each parameter layer was prepared, classified, rated, and weighted according to the conventional DRASTIC framework. The initial vulnerability map was then produced by overlaying the weighted layers.
To assess temporal uncertainty, two dynamic parameters, including depth to groundwater and net recharge, were analyzed over a 20-year period. For each parameter, three statistical states were defined: Q25, median, and Q75. The median represented the average condition, while Q25 and Q75 represented dry and wet conditions, respectively. The median rainfall of the study period was calculated as 680 mm, and recharge was estimated based on rainfall and hydrometric discharge data.
After generating vulnerability maps for the three temporal states, the DRASTIC index was calibrated using nitrate concentration. Two hybrid models were applied: RF-HHO as a machine learning approach and CNN-HHO as a deep learning approach. In both models, DRASTIC weights and ratings were considered as decision variables. The weight range was set from 1 to 7, and the rating range was set from 1 to 10. The objective function was defined as maximizing the correlation between the calibrated vulnerability index and nitrate concentration. Finally, the performance of the classical and calibrated DRASTIC models was compared using the correlation coefficient.
The results of temporal uncertainty analysis showed that the two dynamic DRASTIC parameters, namely depth to groundwater and aquifer recharge, had noticeable temporal variability over the 20-year period. For groundwater depth, the analysis indicated that changes in groundwater level affected the rating classes of large parts of the aquifer. Under wet conditions, groundwater depth changed by approximately 20 to 50 cm in different parts of the aquifer, whereas under dry conditions, the magnitude of change ranged from about 30 to 70 cm. The largest changes in rating distribution were observed in the dominant rating classes 7 and 9, which covered a large portion of the aquifer. However, because these dominant classes are close in terms of DRASTIC rating, the influence of groundwater depth variability on the final vulnerability index was less pronounced than expected from its coefficient of variation.
The analysis of recharge showed that recharge variability had a stronger effect on the final vulnerability zoning than groundwater depth, despite having a lower coefficient of variation. The coefficient of variation of recharge under dry conditions was higher than that under wet conditions, indicating that recharge is more sensitive to rainfall reduction in dry periods. The vulnerability maps generated under the three temporal states demonstrated that recharge plays a key role in controlling the vulnerability pattern of the Sari–Neka aquifer. In the Q25 scenario, representing drier conditions and lower recharge, a larger area of the aquifer was classified in the very low vulnerability class. In contrast, in the Q75 scenario, representing wetter conditions and higher recharge, the area of high vulnerability classes increased. This indicates a direct relationship between recharge and the DRASTIC vulnerability index. Higher recharge increases the potential for pollutant transport from the land surface to groundwater and consequently increases the vulnerability score.
The comparison of the three DRASTIC vulnerability maps confirmed that temporal uncertainty can shift the spatial extent of vulnerability classes. The dry scenario reduced the area of high vulnerability zones, while the wet scenario expanded them. This finding is particularly important for coastal and agricultural aquifers because pollutant migration risk may increase during wetter periods when recharge is higher. Therefore, a single static vulnerability map based on long-term mean values may underestimate or overestimate vulnerability depending on the hydrological condition considered.
The calibration results showed that the classical DRASTIC model had a relatively weak correlation with nitrate concentration. The correlation coefficient between the initial DRASTIC index and nitrate concentration was 0.37. This result indicates that the standard weights and ratings of DRASTIC are not sufficiently adapted to the local conditions of the Sari–Neka aquifer and have limited ability to represent nitrate-related vulnerability.
After calibration with the RF-HHO model, the correlation coefficient increased to the range of 0.62 to 0.64. This improvement shows that the optimization of DRASTIC weights and ratings using a machine learning-based approach can significantly enhance the relationship between the vulnerability index and nitrate concentration. The RF-HHO model was able to improve the performance of the DRASTIC index by identifying more appropriate parameter contributions under local hydrogeological conditions.
The CNN-HHO model provided the best calibration performance. The correlation coefficient between the CNN-HHO calibrated DRASTIC index and nitrate concentration increased to the range of 0.67 to 0.68. The higher performance of CNN-HHO compared with RF-HHO suggests that deep learning is more effective in identifying nonlinear and spatial relationships between DRASTIC parameters and nitrate concentration. This is particularly relevant in groundwater vulnerability assessment, where vulnerability patterns are spatially structured and affected by interactions among hydrogeological, climatic, and land surface factors.
The analysis of optimized weights and ratings also provided important insights. The greatest variation in optimized weights was observed for the aquifer media parameter, while the lowest variation was observed for topography. In terms of final optimized weights, depth to groundwater received the highest weight, indicating its importance in controlling the potential for contaminant movement to groundwater. Topography received the lowest weight, suggesting that slope had a relatively limited role in explaining nitrate distribution in the Sari–Neka aquifer compared with other parameters. These results confirm that the sensitivity of the final DRASTIC index depends not only on the magnitude of temporal variation in each parameter but also on parameter weight, rating structure, spatial distribution, and interaction with other parameters in the index.
This study demonstrated that temporal uncertainty in dynamic DRASTIC parameters can significantly influence groundwater vulnerability assessment. The results showed that using fixed long-term values for groundwater depth and recharge may simplify the actual behavior of the aquifer and may not fully represent vulnerability under dry and wet conditions. The comparison of Q25, median, and Q75 scenarios revealed that drier conditions increased the extent of very low vulnerability areas, while wetter conditions expanded high vulnerability zones. This finding highlights the importance of recharge in controlling pollutant transport potential in the Sari–Neka coastal aquifer.
The study also showed that a larger coefficient of variation in a parameter does not necessarily result in a stronger effect on the final vulnerability map. Although groundwater depth had a higher coefficient of variation than recharge, its effect on the final vulnerability index was less visible due to the dominance of close rating classes, particularly ratings 7 and 9. In contrast, recharge had a stronger influence on vulnerability zoning despite its lower coefficient of variation. Therefore, the sensitivity of the DRASTIC index should be interpreted based on the combined effects of weight, rating, spatial distribution, and parameter interaction.
The calibration results confirmed that data-driven optimization can substantially improve the performance of the DRASTIC model. The correlation between the classical DRASTIC index and nitrate concentration was 0.37, while calibration increased it to 0.62–0.64 using RF-HHO and 0.67–0.68 using CNN-HHO. The CNN-HHO model showed the best performance, indicating the advantage of deep learning in extracting spatial and nonlinear relationships between vulnerability parameters and nitrate contamination.
Overall, the integration of temporal uncertainty analysis and data-driven calibration provides a more realistic framework for groundwater vulnerability assessment. The calibrated DRASTIC model, particularly when combined with CNN-HHO, can be considered a more effective tool for identifying vulnerable zones in coastal and agricultural aquifers. For future studies, it is recommended to incorporate additional anthropogenic variables such as land use, fertilizer application, irrigation intensity, well density, and pollution source distribution to further improve the prediction of nitrate vulnerability.
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
“Conceptualization, Amirhossein Kiarazm. Seyedeh Hoda Rahmati. and Hossein Babazadeh. Aminreza Neshat.; methodology, Amirhossein Kiarazm. Seyedeh Hoda Rahmati.; software, Aminreza Neshat.; validation, Hossein Babazadeh., Seyedeh Hoda Rahmati. and Amirhossein Kiarazm.; formal analysis, Hossein Babazadeh.; investigation, Amirhossein Kiarazm.; resources, Seyedeh Hoda Rahmati.; data curation, Seyedeh Hoda Rahmati.; writing—original draft preparation, Amirhossein Kiarazm.; writing—review and editing, Amirhossein Kiarazm. Seyedeh Hoda Rahmati.; visualization, Hossein Babazadeh.; supervision, Seyedeh Hoda Rahmati.; project administration, Aminreza Neshat.; All authors have read and agreed to the published version of the manuscript.”
This declaration does not apply to the use of basic tools for checking grammar, spelling, references, etc. If there is nothing to disclose, there is no need to add a statement.
The datasets generated and/or analyzed during the current study are not publicly available due to data ownership and confidentiality restrictions but are available from the corresponding author upon reasonable request and with permission from the data owner.
The author gratefully acknowledges the valuable support and cooperation of the Water Studies Office of the Water Research Institute, Ministry of Energy, throughout the development and preparation of this manuscript.
The authors avoided data fabrication, falsification, and plagiarism, and any form of misconduct.
The authors declare no conflict of interest.
