Spectral reconstruction and variety identification of brewing sorghum based on spectral adaptive feature enhancement network

Accurate identification of brewing sorghum varieties is the key to guaranteeing the stability of liquor quality. In response to the problem that the detail loss and band distortion caused by imaging system noise lead to a decrease in spectral reconstruction accuracy, this study proposes a detection algorithm based on the Spectral Adaptive Feature Enhancement-based Multi-stage Spectral-wise Transformer (ASTE-MST++) network —an improved version of the original Multi-stage Spectral-wise Transformer (MST++) network, which integrates a self-designed spectral adaptive feature enhancement (ASTE) module. This module adopts dynamic threshold denoising and frequency domain attention enhancement to enhance the local detailed features in the image, thereby optimizing spectral reconstruction quality. To evaluate the spectral reconstruction performance, the proposed spectral reconstruction model is compared with the existing Hyperspectral Convolutional Neural Network-Dense (HSCNN-D) and the MST++ network. Subsequently, classification models including Partial Least Squares Discriminant Analysis (PLS-DA), Support Vector Machine (SVM), Random Forest (RF), and Sparrow Search Algorithm-Optimized Random Forest (SSA-RF) are constructed to test the reconstructed data. The results show that the ASTE-MST++ model integrated with SSA-RF achieves the best performance in sorghum variety detection (accuracy 92.66%, recall 92.31%, F1-score 92.12%). The proposed ASTE-MST++ detection model provides an efficient and practical solution for the online sorting of brewing sorghum. • Efficient spectral reconstruction of sorghum varieties using the ASTE-MST++ network. • An ASTE module is proposed to enhance spectral reconstruction network performance. • The SSA-RF model applied on the reconstructed spectra showed impressive performance. • The proposed ASTE‑MST++ model surpasses MST++ in spectral reconstruction quality.