Prediction of milk fat content based on deep learning combined with hyperspectral imaging techniques

Fat content is one of the key indicators for evaluating the quality of milk. Hyperspectral imaging (HSI) can effectively capture changes in the surface and internal composition of milk. However, traditional data analysis methods lack the ability to model the nonlinear relationship between spectral information and fat content. This study proposes a joint learning-based selector and predictor (JLSP) model based on deep learning —a method capable of simultaneously performing feature selection and model construction. Inspired by reinforcement learning, JLSP employs the actor-critic algorithm to optimize the feature selection strategy. In this study, HSI data of milk covering its entire shelf life were collected, and fat content was measured. Preprocessing methods were applied to eliminate interference from lighting and environmental factors in milk spectral data. JLSP was utilized for feature wavelength selection and fat content prediction, and its performance was compared with conventional regression models, feature selection methods and deep learning models. The JLSP shows the best performance on the dataset, achieving a test set R 2 of 0.9734 and a mean square error (MSE) of 0.0573. Subsequent spectral analysis confirmed the interpretability of JLSP’s feature selection results, further validating the feasibility of using deep learning methods to predict fat content in milk hyperspectral datasets. • Hyperspectral imaging technology integrated with deep learning facilitates non-destructive detection of milk fat content. • A dataset comprising 1245 hyperspectral images alongside corresponding physicochemical parameters was constructed from 83 types of milk. • A deep learning framework capable of simultaneous feature selection and regression prediction was proposed. • A joint learning strategy was proposed to achieve collaborative training of multiple neural networks.