Raman spectroscopy for multi-label identification of common apple pesticide mixtures using CNNs and gradient-weighted class activation mapping

[EN] Fruits on the market must be free of harmful pesticides, but residues often persist, posing a risk to public health. Detecting these residues efficiently is crucial. However, traditional methods are slow and lack real-time capability. Besides, the detection challenge becomes even more difficult when pesticides are used in mixtures, and Raman spectroscopy offers a promising real-time solution for identifying pesticide residues. This study investigates the detection of commonly used pesticide active compounds on apples in complex mixtures via Raman spectroscopy and a multilabel classification approach, addressing the challenge of identifying active compounds within complex mixtures. Models were also tested on independent, entirely different mixtures to assess their limitations and robustness. Unique spectral fingerprints were established for pure active compounds, which enabled differentiation in mixture analyses. Dimensionality reduction techniques (t-SNE and PCA) effectively distinguished pure compounds from mixtures. Three machine learning models: Partial Least Squares (PLS), Support Vector Machine (SVM), and a 1-D Convolutional Neural Network (1-D CNN) were trained to identify active compounds in mixtures. All models classified key compounds with high F1 scores, with Captan detected with F1 scores higher than 99 %, but some compounds, such as Folpet or Mancozeb, showed poor predictions. The 1-D CNN achieved the best performance with the lowest Hamming Loss, and Grad-CAM analysis confirmed it identified relevant spectral regions. This work demonstrates the potential of machine learning-enhanced Raman spectroscopy for effective pesticide monitoring, regulatory compliance, and food safety.