Addressing SERS data scarcity for sulfonamide detection in water using a supervised VAE-WGAN framework

• A supervised VAE-WGAN framework is proposed for SERS data augmentation • Synthetic SERS spectra enable improved sulfonamide classification in water • Specificity increased from 73.33% to 90.00% • Generated data preserves spectral structure and enhances multiple ML models • The approach supports reliable SERS-based environmental monitoring Antibiotic residues, particularly sulfonamides, are increasingly ubiquitous throughout aquatic environments, raising significant concerns across human, animal, and ecosystem health. Surface-enhanced Raman scattering (SERS) spectroscopy offers a rapid and sensitive method for detecting trace amounts of these contaminants, however the effectiveness of machine learning (ML) models for SERS spectra analysis is often hindered by data scarcity and low signal-to-noise ratios, especially near detection limits. To address this challenge, a supervised deep learning framework that integrates a Variational Autoencoder and a Wasserstein Generative Adversarial Network (VAE-WGAN) with a Multilayer Perceptron (MLP) classifier was developed. The model was applied to the identification of three sulfonamide antibiotics (sulfamethoxazole, sulfapyridine and sulfathiazole) in tap water using a COF–Au nanoparticle composite as SERS substrate and a portable Raman system. The generative module produces realistic synthetic SERS spectra, enabling effective data augmentation and joint optimisation of the classification task. Compared to a standard MLP trained solely on experimental data (78.77% accuracy), the proposed MLP-VAE-WGAN model demonstrated significantly higher classification performance (82.19% accuracy) and, importantly, a substantial increase in specificity (from 73.33% to 90.00%), indicating an improved ability to reject non-target samples in a multi-class setting. The generated data not only preserved class structure but also improved performance across several conventional ML classifiers and outperformed conventional data augmentation techniques. These findings highlight the potential of generative models as a scalable tool for improving classification tasks in data-limited spectroscopic applications towards environmental monitoring.