Feasibility Study on Identifying Seed Variety of Soybean With Hyperspectral Imaging and Deep Learning

Abstract

Seed variety purity is an important indicator of seed quality, and mixing soybean seeds at different maturity stages can affect crop growth and food quality. This study investigated the feasibility of recognizing five soybean varieties at different maturity stages using hyperspectral imaging. Hyperspectral data from 3600 soybean seeds were collected in the range of 395.5–1003.7 nm. First, the potential to qualitatively distinguish the five soybean varieties was assessed using visual cluster analyses based on principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and uniform manifold approximation and projection (UMAP). Next, the performance of four classification models—random forest (RF), extreme learning machine (ELM), partial least squares discriminant analysis (PLS-DA), and one-dimensional convolutional neural network (1DCNN)—was compared. Multiplicative scatter correction (MSC) preprocessing significantly improved the recognition effect of all four models, with the 1DCNN model demonstrating the highest accuracy and most stable recognition performance. The effects of feature bands extracted using competitive adaptive reweighted sampling (CARS), variable importance in projection (VIP), and local linear embedding (LLE) on the four models were also compared. The accuracy of all four feature band sets, when combined with the MSC+1DCNN model, exceeded 96% in identifying soybean varieties. Therefore, these results indicate that the 1DCNN discriminant analysis model is suitable for spectral data analysis in soybean seed variety classification and can significantly enhance classification accuracy.