Efficient wheat variety identification using Raman hyperspectral imaging in combination with deep learning

Abstract

Wheat (Triticum aestivum L.) is recognized as a globally important staple crop, with its varietal differences influencing food processing, nutritional value, and agricultural productivity. Traditional identification methods are often considered inefficient and subjective, while existing spectral techniques are hindered by complex preprocessing procedures and limited model interpretability. To address these limitations, an efficient and interpretable approach was developed by integrating Raman hyperspectral imaging with deep learning techniques. First, a segmentation framework, One-Target Hyperspectral Image Segmentation and Extraction based on the Segment Anything Model, was developed to efficiently and reliably extract regions of interest from wheat grains in Raman hyperspectral images. Subsequently, Raman characteristic peaks were selected using chemical prior knowledge, rather than traditional data-driven methods that rely on statistical features, to enhance the chemical interpretability of the features. Finally, a Raman Spectral Attention Network was designed by incorporating multiscale feature extraction and a Transformer module to improve the modeling performance on the selected Raman characteristic peaks. Experimental results demonstrated that the segmentation framework significantly improved preprocessing efficiency, while Raman Spectral Attention Network achieved an accuracy of up to 99 % in classifying eight wheat varieties. Overall, this study provides a reliable, interpretable, and efficient solution for wheat variety identification, with promising applications in food quality assessment, precision agriculture, and food safety monitoring.