Spectral–spatial fusion attention transformer for brewing wheat variety identification and food-processing raw-material quality evaluation

Abstract

Current spectral reconstruction methods exhibit limitations in jointly modeling spatial and spectral features, making it difficult to fully capture complex spectral information and thereby constraining reconstruction fidelity. To address this challenge, this paper proposes a novel Multi-stage Spectral–Spatial Fusion Attention Transformer (MSFAT), designed to enable efficient fusion and collaborative modeling of spatial-spectral data, ultimately improving the accuracy of spectral reconstruction. The core innovation lies in the development of the Spectral–Spatial Fusion Attention Module (SFAM), which integrates a dual frequency-spatial attention mechanism. By combining fast-Fourier transform with channel attention, SFAM effectively extracts global spectral dependencies and local spatial features synergistically. This mechanism allows SFAM to capture long-range spectral correlations in the frequency domain while emphasizing critical spatial details, significantly boosting the model's capacity to handle complex hyperspectral data. Experimental results validate the efficacy of MSFAT, showing a 0.37 dB improvement in PSNR over MST++ for hyperspectral image reconstruction, along with a 5.6 % reduction in MRAE and a 4.8 % decrease in RMSE. Furthermore, a one-dimensional (1D)-CNN-based brewing wheat variety classification model was developed using the reconstructed hyperspectral imaging data from MSFAT. This model achieved a test accuracy, recall, and F1 score of 96.66 %, 96.66 %, and 96.63 %, respectively—representing a 2.08 % accuracy improvement over MST++ and closely approximating the performance obtained using original spectral data (97.08 %). These findings demonstrate that integrating MSFAT with 1D-CNN offers a high-precision and cost-effective solution for the identification of brewing wheat raw materials, underscoring its broad potential for applications in smart agriculture and food composition analysis.In addition, the spectral data reconstructed by the proposed model can be further applied to analyze physicochemical characteristics relevant to food processing, such as starch, protein, and moisture content. This RGB-based spectral reconstruction approach provides a new avenue for raw-material sorting, quality assessment, and process monitoring in brewing and other food-processing operations.