Leveraging SERS and\nTransformer Models for Simultaneous\nDetection of Multiple Pesticides in Fresh Produce

The widespread use of pesticides in agriculture poses\nfood safety\nand environmental risks, highlighting the need for rapid detection\ntechniques to mitigate contamination. Surface-enhanced Raman spectroscopy\n(SERS) coupled with machine learning provides a powerful approach\nfor the detection and quantification of multiple pesticides in agricultural\nproducts. This study introduces the SERSFormer-2.0 model, which excels\nin both multilabel classification and multiregression tasks for pesticide\nanalysis, leveraging the power of transformer-based machine learning\narchitectures. SERSFormer-2.0 employs novel multitask learning approach\nwith task specific feature representation layers, shared multihead\nattention transformer encoder, and task-specific output layers to\ndetect pesticides and estimate the precise concentrations of each\npesticide simultaneously. By utilizing core–shell gold–silver\nnanoparticles, the model achieves near-perfect performance in identifying\nand quantifying pesticide residues, with multilabel metrics and regression\naccuracy demonstrating exceptional reliability (accuracy = 0.999;\nF1 score = 0.992; precision = 0.990; recall = 0.996). A detailed examination\nof the Raman spectra reveals the predominant influence of certain\npesticides, and the mechanisms behind spectral dominance were elucidated.\nOur findings underscore the SERSFormer-2.0 model ’s robustness\nand its potential to detect mixed contaminants in agricultural products,\nenhancing food safety and regulatory practices.