In order to establish a rapid and nondestructive method for the detection of protein in walnut kernel, a prediction model for walnut protein content was established based on near infrared spectroscopy. Near infrared band selection method was also studied for optimizing protein content predicting model. Firstly, near infrared spectroscopy information of three different sized walnut samples in the full range of 1040~2560nm was collected, and the pretreatment was completed by multiple scattering correction and standard normal method. Then the spectral characteristic bands were selected by interval partial least square algorithm，and the partial least squares prediction models were established at full bands ranging from 1040nm to 2560nm and selected feature bands, respectively. The spectral analysis results showed that the walnut size has no significant effect on the walnut protein content prediction using near infrared spectrum method. The root mean square error and the correlation coefficients of the optimized model were 0.021 and 0.913 in the whole walnut sample validation set, which indicated that the application of interval partial least algorithm could optimize the model quality and reduce the complexity.