The limited nutritional information provided by external food representations has constrained the further development of food nutrition estimation. Near-infrared hyperspectral imaging (NIR-HSI) technology can capture food chemical characteristics directly related to nutrition and is widely used in food science. However, conventional data analysis methods may lack the capability of modeling complex nonlinear relations between spectral information and nutrition content. Therefore, we initiated this study to explore the feasibility of integrating deep learning with NIR-HSI for food nutrition estimation. Inspired by reinforcement learning, we proposed OptmWave, an approach that can perform modeling and wavelength selection simultaneously. It achieved the highest accuracy on our constructed scrambled eggs with tomatoes dataset, with a determination coefficient of 0.9913 and a root mean square error (RMSE) of 0.3548. The interpretability of our selection results was confirmed through spectral analysis, validating the feasibility of deep learning-based NIR-HSI in food nutrition estimation.

The total viable count (TVC) is an important index that reflects the freshness, contamination, and spoilage of dishes. Presently, near-infrared spectroscopy (NIR) is widely used for the assessment of TVC in raw meat, but there are few reports on the application of NIR spectroscopy to evaluate the TVC of Chinese meat dishes. And in the existing research, the modeling samples are industrial products, which makes it lack extensive representation, resulting in insufficient generalization ability of the model. Therefore, we prepared a relatively complex modeling sample to improve the application ability of the model. In this research, braised beef was investigated as an example to establish a predictive model of TVC in complex food matrices using NIR spectroscopy. For this purpose, five groups of 472 braised beef samples with different seasonings were used to collect NIR spectra (900-1700 nm). For five groups, the periods of storage were tested using spectroscopy every day, after which the TVC was measured directly. Furthermore, we established a PLS prediction model. Later, the model established by using the best pre-processed spectral data for band selection, and the model established by band selection followed by variable processing methods were compared. The model using SPA band selection method followed by median center processing exhibited a good prediction ability, with the coefficient of correlation (R) of 0.8507, root mean squared error of prediction (RMSEP) of 0.2707 lg CFU/g, ratio of prediction to deviation (RPD) of 2.00, and ratio error range (RER) of 9.46. These results prove that the developed model can be used in the prediction of TVC in Chinese braised beef.