Data augmentation for machine learning assisted pesticide detection from SERS.

In Surface-Enhanced Raman Spectroscopy (SERS) based pesticide sensing, a key challenge is the detection, especially at low concentrations, where characteristic Raman vibrational features are often obscured by background or remain unresolved. This creates a needle-in-a-haystack problem, necessitating the use of machine learning (ML) algorithms. However, effective ML training is hindered by the limited availability of diverse spectral data, restricting the model's ability to learn complex patterns. To overcome this challenge we use a transformer-based data synthesizer named TabuLa to generate high-quality synthetic spectral data that mimics real pesticide signals. By augmenting real data with synthetic samples, we can enhance dataset diversity, improving the robustness and generalization capabilities of ML models. The performance of TabuLa was evaluated by comparing the similarity of real and synthetic datasets and by testing the ability of supervised ML models trained on synthetic data to detect the presence of pesticides from real data. Our results demonstrate that TabuLa can generate realistic synthetic data and significantly improve ML performance in pesticide detection, offering a promising solution to data limitations in SERS applications.