Detection of Rice Prolamin and Glutelin Content Using Hyperspectral Imaging Combined with Feature Selection Algorithms and Multivariate Regression Models.

Abstract

Prolamin and glutelin are the major constituents of rice protein. The rapid and non-destructive detection of prolamin and glutelin content is conducive to the accurate assessment of rice quality. In this study, hyperspectral imaging combined with regression models and feature wavelength selection was employed to detect the rice prolamin and glutelin content. Feature wavelength selection was achieved using the successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), and convolutional neural network (CNN)-based Gradient-weighted Class Activation Mapping++ (GradCAM++). Partial least squares regression (PLSR), support vector regression (SVR), back-propagation neural network (BPNN), and CNN models were established using the full spectra and the feature wavelengths. The BPNN models showed the best prediction performance for prolamin and glutelin. The optimal BPNN models achieved a correlation coefficient ( r ) greater than 0.8 for both proteins. Performance differences were observed between models using feature wavelengths and those using the full spectra. The GradCAM++ method was used to select feature wavelengths with different threshold values, and the performance of different threshold values were compared. The results demonstrated that hyperspectral imaging with multivariate data analysis was feasible for predicting the rice prolamin and glutelin content. This study provided a methodological reference for detecting prolamin and glutelin in rice, as well as the other protein types.

Plant phenotyping relevance

イネ種子のタンパク質含量という植物形質を、ハイパースペクトル画像と波長選択・回帰モデルで非破壊推定する手法が研究の中心であるため。

Code and data availability

The article describes hyperspectral imaging of rice with PLSR/SVR/BPNN/CNN modeling, but contains no public dataset deposit, no author code/model availability statement, no supplementary files (pmc-prop-has-supplement no), and no repository identifiers. Only ORCID profiles and the CC BY license URL appear; none are a p

No evidence-backed public reproduction asset is currently recorded.