Abstract: In order to detect the water content in camellia seeds quickly and accurately, and also to solve the problems of time-consuming and labor-intensive of traditional drying and detection methods, a non-destructive test method for water content in camellia seeds was proposed based on hyperspectral technology. Camellia seeds were selected as the research object, the water content in camellia seeds was detected, and spectral models were established. The camellia seeds spectrum was pretreated with Savitzky-Golay (SG) convolution smoothing, first-order differential, second-order differential, and multiple scattering correction (MSC), respectively, and effective sensitivity wavelengths were extracted through stepwise regression. Partial least squares regression (PLSR), back propagation (BP) neural network, and radial basis function (RBF) neural network were used to establish prediction models. External verification was conducted for the established models, and the optimal prediction model was selected. It was shown that the spectrally sensitive bands with high correlation coefficients were 410-450, 600-620, 780-880, 940-971 nm, respectively. For the established PLSR model based on spectrum pretreated with MSC, the correlation coefficient and root mean square error were 0.953 4 and 0.22%, respectively, on correction set, and were 0.939 9 and 0.27%, respectively, on validation set, which were higher than those of the established BP neural network and RBF neural network models. Therefore, it was feasible to detect water content in camellia seeds by hyperspectral technology, and the present study could provide basis for the non-destructive online detection of water content in camellia seeds.

Key words: hyperspectral, camellia seeds, water content, spectral analysis, data processing, nondestructive test