A One-Class Anomaly Detection Framework for Groundwater Quality Assessment

Assessing groundwater quality is crucial for public health but is challenged by the complexity of contaminants and the scarcity of labeled data. Traditional methods often require well-defined classes for both normal and anomalous samples, a condition rarely met in practice. This paper introduces a novel evaluation framework using one-class anomaly detection, trained exclusively on a small set of expert-identified, high-quality groundwater samples. We construct an ensemble of One-Class SVM, Isolation Forest, and Local Outlier Factor models and propose a new standard that classifies unseen samples into multiple levels of deviation based on the score distribution of the positive training data. Our ensemble model, validated against expert evaluations on a real-world hydrochemical dataset, demonstrates high accuracy. This provides a robust, nuanced, and practical solution for groundwater assessment in data-scarce environments, effectively establishing a new data-driven benchmark for water quality.