Hyperspectral imaging combined with deep learning models for the prediction of geographical origin and fungal contamination in millet

Abstract

Millet is one of the major coarse grain crops in China. Its geographical origin and Fusarium fungal contamination with ergosterol and deoxynivalenol have a direct impact on food quality, so the rapid prediction of the geographical origins and fungal toxin contamination is essential for protecting market fairness and consumer rights. In this study, 600 millet samples were collected from twelve production areas in China, and traditional algorithms such as random forest (RF) and support vector machine (SVM) were selected to compare with the deep learning models for the prediction of millet geographical origin and toxin content. This paper firstly develops a deep learning model (wavelet transformation-attention mechanism long short-term memory, WT-ALSTM) by combining hyperspectral imaging to achieve the best prediction effect, the wavelet transformation algorithm effectively eliminates noise in the spectral data, while the attention mechanism module improves the interpretability of the prediction model by selecting spectral feature bands. The integrated model (WT-ALSTM) based on selected feature bands achieves optimal prediction of millet origin, with its accuracy exceeding 99% on both the training and prediction datasets. Meanwhile, it achieves optimal prediction of ergosterol and deoxynivalenol content, with the coefficient of determination values exceeding 0.95 and residual predictive deviation values reaching 3.58 and 3.38 respectively, demonstrating excellent model performance. The above results suggest that the combination of hyperspectral imaging with a deep learning model has great potential for rapid quality assessment of millet. This study provides new technical references for developing portable and rapid hyperspectral imaging inspection technology for on-site assessment of agricultural product quality in the future.