In order to improve the detection accuracy of wheat protein model, a method based on optimized back propagation (BP) neural network was proposed in this paper. Using the wheat spectral data, the spectral data after eliminating the abnormal samples were preprocessed differently to determine the optimal preprocessing method. Dung beetle optimizer (DBO) was used to optimize the initial weights and thresholds of BP neural network, and full-band partial least squares regression (PLSR), DBO-BP and BP prediction models of wheat protein content were established. The results showed that Standard Normal Variate Transform (SNV) pre-processing method was the best. Under the same training sample, the detection accuracy of the established DBO-BP model is the highest. Compared with PLSR and BP neural network, the prediction set root mean squared error (RMSEP) is reduced by 58.65% and 46.53%, the correlation coefficient is increased by 10.35% and 5.06%, and the residual prediction deviation (RPD) is 7.27. Therefore, the DBO-BP neural network model can quickly and accurately detect the wheat protein content, which provides a theoretical basis for the detection of wheat protein based on BP neural network.