Machine Learning Meets Hydrogeology: A K-Means Approach to Aquifer Vulnerability Mapping

Abstract

Aquifer vulnerability mapping is essential for groundwater management and pollution prevention. This study compares the traditional DRASTIC method with the K-means clustering algorithm to evaluate the vulnerability of the Quíbor Valley aquifer in northwest Venezuela. The DRASTIC method estimates vulnerability using a weighted sum of seven parameters: water table depth, recharge, aquifer media, soil type, slope, impact of the vadose zone, and hydraulic conductivity. However, its results depend on subjective weight and rating assignments. In contrast, K-means clustering is an unsupervised machine-learning method that groups data based on natural similarities, providing a more objective and data-driven approach. Using Geographic Information System (GIS), thematic maps for each parameter were generated and normalized. K-means identified three clusters: K1 (low vulnerability), influenced by deep water levels, impermeable soils, and low hydraulic conductivity; K3 (high vulnerability), associated with shallow water levels, semipermeable soils, low slopes, and high hydraulic conductivity; and K2 (moderate vulnerability), representing intermediate conditions. The aquifer media, composed mainly of permeable sands and gravels, showed no influence on cluster separation. Comparison of vulnerability maps indicates that both methods outline different vulnerable zones, but K-means reduces the subjectivity of DRASTIC and provides a simpler, robust alternative. Validation using 35 groundwater electrical conductivity measurements showed a positive correlation for K-means (rₒ = 0.69), while DRASTIC produced a negative correlation (rₒ = −0.55). The weaker DRASTIC performance is linked to its emphasis on water table depth and recharge, which have limited variability in this overexploited aquifer.