In order to eliminate the influence of scattering, the original spectrum was processed by multiplicative scatter correction (MSC). The effective band was selected according to the correlation coefficient method, and 8 characteristic wavelengths were selected by continuous projection algorithm combined with information entropy. Finally, the effective bands and different features were used to establish prediction model for mildew maize aflatoxin B₁ and gibberellin content at wavelength by BP neural network. The results showed that the prediction model established by spectral information at 8 kinds of characteristic wavelengths was the best, with the correct prediction rate of aflatoxin B₁ content of 98.74%, the root mean square error of 0.048 5, and the correct rate of gibberellin content prediction of 100%, and the square root error of 0.160 5. Therefore, the method of hyperspectral coupled with neural network is feasible to detect the aflatoxin B₁ and gibberellin content in moldy maize.