Abstract:

To investigate the feasibility of using hyperspectral imaging technique to detect the soluble solid content (SSC)
of postharvest kiwifruits based on the obtained reflectance spectra over the range of 900–1 700 nm, SSC prediction models
were established using partial least squares, support victor machine and back propagation artificial neural networks. The
effects of different input variables on model performance were compared comprehensively at 226 wavelengths in full
spectra. The input variables investigated included 12 and 128 effective wavelengths selected by successive projection
algorithm and uninformative variable elimination, respectively. The results showed that successive projection algorithm
could extract the effective wavelengths efficiently, and it had obvious predominance in simplifying SSC prediction model.
BP neural network had better SSC predication performance. BP network combined with successive projection algorithm had
the best SSC prediction performance with correlation coefficient of 0.924 and root-mean-square error of 0.766 for prediction
set. The present study indicated that hyperspectral imaging technique could be used to detect SSC of postharvest kiwifruits
nondestructively, and the technique is feasible for industrial grading of kiwifruits based on internal quality.

Key words: kiwifruit, hyperspectral image, soluble solid content, BP network, successive projection algorithm