Intelligent detection of seed protein of snap bean based on hyperspectral imaging

This study employed hyperspectral imaging (384–1022 nm) for the non-destructive prediction of protein content in snap bean seeds. After acquiring images of 200 seed samples, exploratory data analysis (EDA) was applied to average spectra to remove highly correlated wavelengths, followed by principal component analysis (PCA) to identify salient wavelengths for modeling. Two distinct feature extraction approaches were implemented: a hybrid method integrating EDA, PCA, and statistical feature extraction (SFE), and a second approach based on k-nearest neighbors mutual information (KNN-MI). The extracted features were used to construct three regression models: K-nearest neighbors (KNN), support vector regression (SVR), and decision tree regression. The KNN-MI features yielded the best performance with the decision tree model (validation R² = 0.8917, RMSE = 0.2280), whereas the EDA–PCA–SFE combination performed optimally with SVR (validation R² = 0.9762, RMSE = 0.2330). The results confirm the efficacy of feature extraction methods coupled with hyperspectral imaging for rapid and accurate protein content prediction in snap bean seeds. • Use 384–1022 nm hyperspectral imaging and two methods to detect snap bean seed protein. • SVR with SFE shows top performance (R²=0.9762, RMSE=0.2330) for protein prediction. • Our approach enables non-destructive, rapid detection vs traditional Kjeldahl method. • This study supports online snap bean seed protein detection system development.