Overcoming the bran layer barrier: predicting milled rice starch pasting properties directly from brown rice hyperspectral images

This work proposed an innovative solution to estimate the starch pasting properties in milled rice directly based on the hyperspectral images (HSI) of brown rice. Spectral interferences caused by bran layers between brown and milled rice were firstly identified at 450–480 nm and 1200–1471 nm. Notably, the domain-adversarial neural network and maximum mean discrepancy (DANN-MMD) were efficiently applied to correct the spectral interferences of bran layer from brown rice, in order to predict their pasting properties measured by rapid visco analyzer (RVA). The convolutional neural network models based on DANN-MMD transferred HSI spectrum of brown rice showed superior predictive performance for their starch pasting properties when using the short-wave infrared (SWIR, 980–2000 nm) data, with RPD values of 2.49 and 3.20 for consistency viscosity and peak time, respectively. Consequently, the integrated framework that combines DANN-MMD with deep learning offers a promising strategy of using HSI technique to estimate starch pasting properties directly from brown rice. • Bran layer caused spectral interferences mainly at 450–480 nm and 1200–1471 nm. • DANN-MMD corrects spectral interferences of bran layer from brown to milled rice. • CNN models had best prediction performance for starch pasting properties in rice. • HSI technique based on DANN-MMD-CNN can assess pasting properties from brown rice.