The protein content of milk will affect the quality of milk.The feasibility of predicting the protein content of milk is studied by using the spectral feature information of hyperspectral image.In this paper,a prediction modeling method (CARS-SPA-BP) based on competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) combined with multilayer feedforward neural network (back propagation,BP) is proposed.In the experiment,250 groups of hyperspectral data of five kinds of milk were collected by the visible/near infrared hyperspectral imaging system.Through the experimental comparison,the standardized method was used to preprocess the obtained absorption spectrum,and then the CARS combined with SPA was used to select the characteristic wavelength,18 characteristic wavelengths are obtained.Through experiments,the determination coefficients R2c and R2ｐ of training set and test set of CARS-SPA-BP model reach 0.971 and 0.968 respectively,and the root mean square error of calibration (RMSEC) and root mean square error of prediction (RMSEP) reach 0.033 and 0.034,respectively.It is found that the prediction results of multilayer back propagation (BP) neural network model based on CARS and SPA are not significantly lower than that of full wavelength model,Therefore,the CARS combined with SPA for wavelength screening and BP neural network can basically complete the prediction of milk protein content.In order to verify the prediction ability of CARS-SPA-BP model,the traditional partial least squares regression (PLSR) is used to model under the same data environment.The experimental results show that CARS-SPA-BP has significantly improved R2p and RMSEP compared with PLSR.The results show that CARS-SPA-BP can make full use of the spectral characteristics of milk to achieve high-precision detection of milk protein content.