Abstract: The PLSR prediction model for beef color of Jingyuan yellow cattle was constructed by hyperspectral image. The hyperspectral images of samples were obtained by visible near-infrared hyperspectral imaging system, the spectral information of the interest regions were extracted, and the average spectrums were calculated. The Monte Carlo method was used to eliminate the abnormal samples, then the sample set was divided and the sample data was preprocessed. Results showed that model with the lightness (L^*) pretreated by the Deresolve method performed best while $R_{C}^{2}$ was 0.979 0 and $R_{P}^{2}$ was 0.976 6. Model with the redness (a^*) pretreated by convolution smoothing (Smoothing-SG) method performed best while $R_{C}^{2}$ was 0.807 0 and $R_{P}^{2}$ was 0.915 5. Model with the yellowness (b^*) pretreated by convolution smoothing (Smoothing-SG) method performed best while $R_{C}^{2}$ was 0.931 1 and $R_{P}^{2}$ was 0.950 6. Characteristic wavelengths were extracted by using competitive adaptive re-weighting method (CARS), continuous projection algorithm (SPA) and non-information variable elimination algorithm (UVE) respectively. The partial least squares regression (PLSR) model based on characteristic band was achieved and further optimized. The best prediction model was combined with the spatial depth and stereoscopic degree of the vision to extract and distinguish form and color perception of the sample. Therefore, it was feasible to construct the PLSR model by using hyperspectral imaging based on values of L^*, a^* and b^*, and the theoretical basis for online rapid detection of beef quality was provided.

Key words: Jingyuan yellow cattle, hyperspectral imaging technology, chromaticity values, partial least squares regression