Abstract:Corylus heterophlla Fisch and Corylus mandshurica Maxim were taken as the research objects.After removing the shell，Corylus heterophlla Fisch and Corylus mandshurica Maxim were made into 90 groups and 60 groups respectively，and the original light with the wavelength range of 900 nm-1 700 nm was obtained.At the same time，the true value of protein content in hazelnut was measured by Kjeldahl Nitrogen method.And the influence of pretreatment methods such as 1st derivative（1-der），2nd derivative（2-der），multiplicative scatter correction （MSC）and standard normal variate （SNV）on the model accuracy was compared，and the pretreatment methods suitable for hazelnut spectrum were selected.The feature bands suitable for protein prediction were selected by the method of backward interval partial least squares（BiPLS），which could simplify the model.A prediction model of protein content was established by using the extreme learning machine（ELM）.Finally，it could be concluded that the correlation coefficient R and root mean square error of prediction（RMSEP）of the two hazelnuts were 0.880 6 and 0.599 3，0.882 3 and 0.598 4 respectively，indicated a high model accuracy.