LatentGW-RiskNet: a latent fingerprint-based autoencoder framework for three-class groundwater risk assessment

Abstract

Groundwater contamination is a significant environmental and public health issue in urban areas with industrial, commercial, and residential activities. Accurate assessment of water quality is challenging due to multiple physicochemical parameters and small sample sizes. It limits the effectiveness of conventional methods for classifying water into Safe, Contaminated, and High-Risk categories. We collected 78 groundwater samples from three functional zones in Rajshahi City. We measured twelve parameters which are pH, turbidity, electrical conductivity, TDS, TSS, TS, COD, Fe, and Mn. We also considered two derived features, MetalRisk and ExceedCount. It reflects metal contamination levels. We proposed LatentGW-RiskNet model along with mitigation strategies to address this issue. A shallow autoencoder based framework that compresses the 10 dimensional input into a 2-D latent fingerprint. It captures water quality information and reduces noise. Latent representations are analyzed using Gaussian Mixture clustering and classification is performed using a Random Forest classifier. The model was evaluated with Leave-One-Out Cross-Validation. It achieved 92% accuracy and F1-score It outperformed with models trained on raw features and standard baselines such as SVM, KNN, and Decision Tree. These results highlight that latent fingerprint representations provide a compact, interpretable, and robust approach for small sample groundwater risk assessment. It enables effective three class classification and supporting informed decision-making for water risk assessment.