Fully non-invasive measurement of protein content in soybean based on spectral characteristics of the pod

Abstract

Cultivation and breeding of legumes as protein sources in the human or animal diet could benefit from accurate, rapid, and non-invasive measurements of protein content. A study was conducted into the feasibility of a fully non-invasive, in vivo protein measurement methodology applied to soybean (Glycine max. L.). The proposed methodology works by recording spectral images of the soybean pods in the visible and near-infrared (Vis-NIR), a rule-based segmentation approach, and partial least squares (PLS) regression to predict the crude protein content of the beans contained within the imaged pods. Using all 150 channels of the spectral camera, a model could be calibrated with a mean absolute precision error (MAPE) of 4.8 % (R2 = 0.92). Applying a tailored feature elimination approach to select only eight spectral bands and degrading the spectral resolution to 25 nm yields a model with a MAPE of 6.0 % (R2 = 0.88), indicating the potential for multispectral cameras in this application.