Design of lightweight neural network model for provenance discrimination of Lycium barbarum based on hyperspectral imaging technology.

The study aimed to build a lightweight neural network model for Lycium barbarum provenance discrimination using hyperspectral imaging technology. Firstly, hyperspectral images (336.2-1038.8 nm) acquisition in 256 bands andregions of interest extraction were carried out. The extracted spectral information was then used to develop two neural network models, i.e. SimpleCNN and FPN, and the discriminative accuracies on the test set reached 90.81 % and 92.74 %, with AUC of 0.9902 and 0.9934, respectively. Subsequently, two lightweight approaches were applied to SimpleCNN. The first approach combined the distillation of attention-shifting knowledge, which increased the accuracy of SimpleCNN to 93.93 % on the test set. As for the second approach using a three-step pruning approach, unstructured pruning removed 96 % of connections (retaining 92.15 % baseline accuracy), while structured pruning reduced neurons or channels by 86 % (preserving 92.30 % accuracy). The lightweight method used in this study avoided the problem of excessive decrease in provenance discrimination caused by feature wavelengths selection. The obtained model has more practical application value in and rapidly discriminating Lycium barbarum provenances.