Raman spectrum classification based on transfer learning by a convolutional neural network: Application to pesticide detection

Abstract

Pesticide detection is of tremendous importance in agriculture, and Raman spectroscopy/Surface-Enhanced Raman Scattering (SERS) has proven extremely effective as a stand-alone method to detect pesticide residues. Machine learning may be able to automate such detection, but conventional algorithms require a complete database of Raman spectra, which is not feasible. To bypass this problem, the present study describes a transfer learning method that improves the algorithm's accuracy and speed to extract features and classify Raman spectra. The transfer learning model described here was developed through the following steps: (1) the classification model was pre-trained using an open-source Raman spectroscopy database; (2) the feature extraction layer was saved after training; and (3) the training model for the Raman spectroscopy database was re-established while using self-tested pesticides and keeping the feature extraction layer unchanged. Three models were evaluated with or without transfer learning: CNN-1D, Resnet-1D, and Inception-1D, and they have improved the accuracy of spectrum classification by 6%, 2%, and 3%, with reduced training time and increased curve smoothness. These results suggest that transfer learning can improve the feature extraction capability and therefore accuracy of Raman spectroscopy models, expanding the range of Raman-based applications where transfer learning model can be used to identify the spectra of different substances.

Keywords: Convolutional Neural Networks; Pesticide residue; Raman spectrum; Sorting algorithms; Transfer learning.