Deep learning-driven multiclass groundwater water quality index assessment using attention mechanisms

Assessing groundwater quality is essential for sustainable water resource management in arid regions, where hydrochemical processes are complex and monitoring resources are often limited. This study proposes a hybrid deep learning framework that integrates convolutional neural networks (CNNs) with transformer-based self-attention mechanisms for multiclass water quality index (WQI) classification. A dataset comprising 621 groundwater samples from the Bechar region in southwestern Algeria was analyzed using 12 hydrochemical parameters. To evaluate model robustness under different monitoring conditions, 12 feature combinations representing distinct hydrochemical groups were tested using both CNN and CNN–Transformer architectures. Model performance was assessed using accuracy, precision, recall, F1-score, the area under the receiver operating characteristic (ROC) curve (AUC), and an error-overlap Venn diagram. The CNN–Transformer consistently outperformed the conventional CNN across most configurations. The highest performance was achieved using the full hydrochemical profile, with an accuracy of 92.5% and a micro-AUC of 0.99. Because the WQI target was deterministically calculated from the same hydrochemical variables used as model inputs, the proposed framework should be interpreted as a surrogate WQI classification model rather than as an independent predictor of groundwater quality. Accordingly, the SHAP importance patterns partly reflect the adopted WQI formulation, including its parameter weights, regulatory limits, and class thresholds. Importantly, several reduced models based on only four to five commonly measured parameters also achieved strong predictive performance, with accuracies of up to approximately 0.85, demonstrating the potential of the proposed framework for streamlined groundwater monitoring when comprehensive laboratory analyses are unavailable.