Identifying the best rice physical form for non-destructive prediction of protein content utilising near-infrared spectroscopy to support digital phenotyping

Abstract Digital rice phenotyping requires rapid assessment of protein content of rice to kernels to support high-throughput crop phenotyping experiments. A fast and non-destructive approach can allow rapid decision making to breed and select relevant rice varieties. Hence, this study compares the predictive potential of near-infrared (NIR) spectroscopy for three physical forms of rice i.e., rice kernel (with glume), whole grain brown rice and powdered rice. The aim is to identify the best physical form to be adapted in future use for high-throughput protein content prediction in rice samples. The models were optimized by selecting key wavelengths most correlated to the protein content in rice. For variable selection, a total of 8 recently developed chemometric variable selection techniques were used. As a baseline comparison to variable selection techniques, partial-least square (PLS) regression analysis was used. The results showed that for all forms of rice samples, variable selection improved the predictive performance compared to the PLS regression modelling. The best accuracies were obtained for the brown rice samples with a prediction error of 0.349 %. Further, this was achieved with only 12 wavelengths compared to the 304 wavelengths available in the original data set. Based on the results, this study indicates that there is no need to grind the rice samples into powder for using NIR spectroscopy. Hence, NIR spectroscopy can directly be used on brown rice samples and can support the rapid assessment of protein content in rice to support digital phenotyping.