Abstract: Near-infrared hyperspectral imaging combined with partial least regression (PLSR) models for rapid and non-destructive detection of beef adulterated with pea protein was investigated. The adulteration samples were prepared by mixing pea protein into minced beef with a concentration gradient of 1%~30% (w/w) (interval of 1%), and a total of 93 samples were finally obtained to collect spectral information. After preprocessed by five methods including moving average smoothing (MAS), Gaussian filter smoothing (GFS), baseline correction (BC), Savitzky Golay convolution smoothing (SGCS), and standard normal variable correction (SNV), the spectra were used to construct prediction models by using PLSR algorithm. The regression coefficient (RC), stepwise and successive projections algorithm (SPA) were then used to select optimal wavelength for model optimization. The results showed that the full-band PLSR model constructed by GFS spectra had better prediction performance (R²_P=0.87, RMSEP=3.22%, ΔE=1.82, RPD=5.82). The 24 optimal wavelengths including 908.8, 913.7, 918.7, 928.5, 936.8, 945.0, 961.5, 971.3, 994.4, 1017.4, 1033.9, 1099.7, 1135.8, 1167.1, 1196.7, 1211.5, 1453.6, 1549.3, 1607.1, 1633.7, 1660.1, 1678.4, 1683.4 and 1686.7 nm were selected by RC method from the GFS spectra and the original PLSR model optimized with these wavelengths showed better performance (R²_P=0.90, RMSEP=2.85%, ΔE=1.13, RPD=6.19). In conclusion, it was possible to use hyperspectral imaging to achieve rapid and non-destructive detection of beef adulterated with pea protein.