Estimation of grain quality parameters in rice for high‐throughput screening with near‐infrared spectroscopy and deep learning

Abstract

AbstractBackground and ObjectivesGrain quality is a complex trait in rice, compared with other staple crops as it is predominantly consumed as a whole grain. Although considered secondary to yield, to align with consumer preferences, breeders are increasingly interested in quality. At the early stages of a breeding program, grain quality‐related traits are often ignored as they are arduous and time‐consuming. Near‐infrared spectroscopy (NIRS) could be a suitable high‐throughput alternative to conventional wet chemistry and image processing‐related methods to be adopted for early screening. This study aims to quantify traits essential for rice breeders such as amylose, chalkiness, length, width, and the length/width ratio in rice samples with NIRS. We used conventional algorithms such as principal component analysis (PCA), partial least square regression (PLSR), multilayer perceptron (MLP), support vector classification (SVC), and linear discriminant analysis (LDA) to compare with the proposed convolutional neural network (CNN) for regression and classification.FindingsOur results showed that the proposed CNN outperformed the conventional models in estimating all traits. Unlike conventional models, CNN models could be developed with raw spectra with minimal to no preprocessing, and along with the transfer‐learning capabilities, the time required for model development could be significantly reduced.ConclusionWe recommend NIRS for quantitative estimation of amylose and chalkiness in rice and rather use classification/categorized estimation for other physical dimension‐related traits such as length and length/width ratio.Significance and NoveltyWe found NIRS to be an appropriate alternative to wet chemistry and image‐based methods for screening lines at the early stages of the breeding program. Estimation of physical parameters such as length and length/width ratio with NIRS is novel and appears reasonable for high‐throughput applications.