Ensemble machine learning enhanced Raman spectroscopy for high accuracy and noise robust classification of carbamate...

The global rise in pesticide contamination demands innovative detection technologies to overcome the limitations of conventional methods. In response to the challenges, this work introduces a machine learning enhanced Raman spectroscopy platform for carbamate pesticides, a class of compounds frequently associated with food safety incidents and ecological contamination. High resolution Raman spectra were systematically collected from four structurally diverse carbamates, and advanced data augmentation techniques were employed to build a robust spectral library, enhancing the model's capacity to manage spectral variability. The classification performance of three machine learning algorithms was compared under different signal-to-noise Ratios (SNR), including Random Forest (RF), Back Propagation Neural Network (BPNN), and Support Vector Machine (SVM). The RF algorithm achieved superior accuracy (98.99%) at 30 dB SNR, whereas BPNN showed stronger performance (65.43%) at 5 dB SNR, and SVM exhibited greater stability under various SNR ratio conditions. To leverage the respective strengths of each algorithm across varying SNR levels, a soft-voting ensemble system was developed. It combines three machine learning classifiers via an optimized fusion algorithm to enhance discriminatory performance across varying SNR. The system demonstrated high classification accuracy, with 98.99% at 30 dB SNR and 65.43% at 5 dB SNR, representing a notable improvement over traditional manual spectral analysis. This work presents a machine learning framework for Raman spectroscopy-based identification of pure carbamate pesticide standards. Further evaluation highlighted the model's proficiency in resolving subtle spectral features associated with carbamate molecular vibrations, even under significant noise interference, establishing a foundational approach for future applications in complex matrices.