Classification of Radish Seeds Using Hyperspectral Imaging and Deep Learning Method

Based on the VIS-NIR hyperspectral imaging technique，a rapid and nondestructive method was investigated for discriminating varieties of radish seeds.After removing noise band，the hyperspectral imaging system with spectrum range of 480.46-1001.6 nm was used to collect six varieties of radish seeds containing 411 bands of hyperspectral images.Savitzky Golay（SG）smooth and multiple scattering correction（MSC）were used to eliminate high frequency superposition of random error.The stack autoencoder（SAE），the successive projections algorithm（SPA）and the variable iterative space shrinkage approach（VISSA）were used to reduce dimensionality of hyperspectral data of radish seeds.Softmax and the support vector machine（SVM）classification model were applied to identify radish seeds samples after dimensionality reduction．Experiment results showed that optimal model was the SAE-Softmax model，and accuracy of training set and accuracy of prediction set by the algorithm were reached 99.72% and 96.22%，respectively.The study demonstrated that VIS-NIR hyperspectral imaging technique was potential for nondestructive classification of radish seed.It was feasible and efficient to apply classification model into seed varieties nondestructive testing analysis.