| In order to explore the probability of nondestructive testing based on hyperspectral image technology, the experiment collected hyperspectral image data of 80 pyrifolia samples range in 400~1 000 nm and its sugar content. Then pretreat the original spectrum data by SNV, MSC, S.Golay, Baseline methods, and found the denoising effect of MSC pretreatment was the best. Then compressed the spectrum data of MSC pretreatment through uninformative variable eliminate method. Finally respectively established the BP neural network and PLS prediction model. The experiment results showed that uninformative variable eliminate had compressed the spectrum variable to 234, which could reduce the modeling of input variable effectively. The PLS prediction model and BP neural network prediction correlation coefficient were both more than 0.85, but PLS prediction model of the correlation coefficient was 0.9508, the root mean square error was 0.268, which was higher than the BP neural network model . |
