Near infrared NIR spectroscopy is a fast, non-destructive quantitative analysis tool that has been widely used in various industries. How to build an effective and accurate model is of importance to the application of NIR spectroscopy. At present, most commonly used quantitative analysis methods are based on shallow models, while Deep Belief Network (DBN) is a probability-based deep model. It can automatically learn the effective feature representation of the input, and as long as the number of last hidden layer output nodes is lower than the dimension of the input spectrum, the spectral data can be reduced in dimension while the feature extraction is completed on the spectral data. Near-infrared spectroscopy is characterized by a large sample size, large variables, and high dimensionality. This paper proposes a near-infrared spectroscopy modeling method based on a deep belief network to estimate the physical concentration. The method uses near-infrared spectroscopy data as the input layer. Firstly, unsupervised learning of the multi-restricted Boltzmenn Machines (RBM) is employed to achieve the feature extraction of the spectrum itself. Then the target physicochemical value is used to fine tune the network, and optimize model parameters. Based on the DBN calibration model, the final regression layer of the deep belief network is developed by the PLS method, and the DBN-PLS calibration model may avoid the optimal local problem caused by the gradient descent algorithm. In this paper, the feasibility of DBN modeling and DBN-PLS modeling is verified by two model evaluation indexes including decision coefficient (R2) and mean square error (mse), and the traditional BP modeling and DBN modeling are compared and analyzed. The analysis results show thatDBN method modeling and DBN-PLS method modeling can improve the prediction accuracy.