Abstract
This study analyses groundwater quality in the Rhône–Mediterranean–Corsica (RMC) basin using 27,741 samples collected from 7,625 sampling points and 224 groundwater bodies, covering a 34-year period (1987–2021). The objective is to quantify, using statistical and machine learning methods, the respective contributions of temporal variability, spatial variability, and lithology to the hydrochemical and microbiological composition of aquifers. The results reveal a marked dichotomy between microbiological and physico-chemical parameters. Faecal contamination (E. coli, enterococci) is dominated by a strong temporal component (33–36% of variance), whereas spatial structuring by lithology remains very limited (< 11%). Conversely, major physico-chemical parameters (calcium, conductivity, sulphates, etc.) exhibit near-zero temporal variability (<7%) and a spatial structuring strongly controlled by lithology, with gains in explained variance of up to 20% through lithological grouping. Linear discriminant analysis achieves an overall classification rate of 56.2%, but with wide disparities: the crystalline basement is perfectly discriminated (76%), while alluvial formations are not discriminated at all (0%), as their chemical signature is more strongly influenced by anthropogenic pressures than by lithology. Machine learning methods, particularly k-nearest neighbours (k-NN), significantly improve performance (76.2% vs. 56.2% for LDA), confirming the existence of nonlinear structures in the data. Unsupervised validation via K-means confirms the reality of the lithological signal. From an operational standpoint, these findings justify a differentiated approach to groundwater management: a geological approach for chemical parameters, and an approach based on land use and agropastoral practices for bacteriological parameters. The proposed methodology – combining PCA, HCA, nested ANOVA, LDA and machine learning – provides a transferable analytical framework applicable to other major European river basins.