Rapid multi-task detection of rice components using convolutional bidirectional long short-term fusion network combined with hyperspectral imaging

Rice is a primary raw material for strong-flavor liquor, and its fat, protein, moisture, and amylose contents directly affect the fermentation efficiency and flavor characteristics of the liquor. This study aimed to simultaneously predict the contents of these four components in rice using hyperspectral imaging (HSI) with different spectral ranges (VIS-NIR: 400–1000 nm and NIR: 900–1700 nm) combined with deep learning methods. A hybrid model (DBiL-Net) of a one-dimensional convolutional neural network (1DCNN) and a bidirectional long short-term memory neural network was developed based on single-task (ST) and multi-task (MT) modeling. Further, the partial least squares regression model preprocessed by MSC-1st andthe 1DCNN deep learning model were used as comparison models. The results showed that the MT DBiL-Net model performed the best within the NIR spectral range. The average accuracy rate Rp² for the MT regression prediction of fat, protein, moisture, and amylose in rice was 0.9703, and the average residual predictive deviation was 9.0218. The results showed that HSI combined with the DBiL-Net model could simultaneously and accurately determine the contents of the four components of rice, thereby providing an efficient method for quality detection of raw materials in liquor.