A hybrid expert–PCA–ANN framework for hydrogeochemical optimization of the water quality index in a semi-arid aquifer

Abstract

Groundwater quality in semi-arid regions is controlled by both natural hydrogeochemical processes and human activities. This study examines and compares four approaches for constructing a groundwater quality index in the southern Chott Hodna Basin (Algeria), combining expert judgment, principal component analysis (PCA), and artificial neural network (ANN) optimization. Eleven physicochemical parameters related to salinity, mineralization, and agricultural contamination were analyzed from 33 groundwater samples. Water quality index values range from 52.67 to 303.44, with a mean of 147.79, which indicates generally poor groundwater quality and marked spatial variability. Higher values are mainly linked to salinity and nitrate enrichment. All WQI weighting approaches show strong agreement (r > 0.97), suggesting consistent representation of the aquifer's hydrochemical structure. The PCA-based approach emphasizes salinity-related variables such as electrical conductivity, total dissolved solids, chloride, and sulfate, which reflect evaporite dissolution and water–rock interaction. In contrast, the ANN-optimized model increases the influence of nitrate, giving prominence to the role of agricultural contamination. The hybrid PCA–ANN model provides the lowest prediction error (around a 12–18% reduction compared to non-optimized models) and better captures complex relationships among variables. Spatial interpolation using ordinary kriging identifies degraded water quality zones mainly in the northern and northeastern parts of the aquifer. In general, combining statistical and machine learning approaches improves the objectivity and reliability of groundwater quality assessment in semi-arid regions.