Improving aquifer vulnerability assessment and its explainability in the Zanjan aquifer: Integrating DRASTIC model and optimized long short-term memory-based metaheuristic algorithms

• Enhanced DRASTIC model using LSTM, PSO, and EO to assess groundwater vulnerability. • LSTM-PSO outperforms other models in predicting nitrate-based groundwater risk. • SHAP explains model outputs, highlighting soil media as the most impactful factor. • Vulnerability maps show highest risk in central aquifer, low nitrate in outer regions. • Coupling deep learning and DRASTIC boosts groundwater management effectiveness by 131 %. Groundwater vulnerability assessment involves the evaluation of the potential contamination risk in specific areas. This study aims to enhance the original DRASTIC model by incorporating three methods to improve its accuracy and reliability in identifying vulnerable groundwater regions (to nitrate) in Zanjan aquifer, Iran. Owing to its remarkable ability, long short-term memory (LSTM) deep learning (DL) algorithm was implemented to predict groundwater vulnerability based on nitrate contamination and DRASTIC framework. The LSTM model was optimized using the particle swarm optimizer (PSO) and equilibrium optimizer (EO) metaheuristic algorithms. The results of the developed models (i.e. LSTM, LSTM-PSO, and LSTM-EO) were compared to those of the original DRASTIC model as a benchmark. Seven parameters of the DRASTIC framework, including depth to groundwater, net recharge, aquifer media, soil media, topography, impact of the vadose zone, and hydraulic conductivity, were used as inputs for the DL models, and the DRASTIC vulnerability index (DVI) derived from the original DRASTIC model, which was adjusted using nitrate samples, was considered as the output. Thereafter, the prioritization of DRASTIC parameters and their effectiveness were evaluated using the mean decrease accuracy (MDA) approach. Additionally, the models explainability was enhanced using the SHapley Additive exPlanations approach. Lastly, all the developed models were evaluated based on the root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R²). The MDA results revealed that the soil media and the impact of the vadose zone parameters exerted the highest and lowest impact on the aquifer vulnerability in the study area, respectively, while the summary plot revealed that soil media exerted the highest impact. The vulnerability maps predicted by the DL model and its optimized versions indicated that the central parts of the aquifer exhibited the highest vulnerability, whereas nitrate concentration was low in the north, northwest, south, southeast, and southwest of the Zanjan aquifer. The correlation coefficient (r) between the nitrate concentrations and the LSTM-PSO, LSTM-EO, LSTM, and DRASTIC models was 0.81, 0.78, 0.76, and 0.35, respectively. This indicates the potential of the PSO and EO algorithms to significantly enhance the performance of the optimized LSTM models compared to the standalone LSTM model. Overall, the coupling of the DRASTIC and deep learning models can be a cost-effective tool in groundwater management.