Estimation of protein content in wheat samples using NIR hyperspectral imaging and 1D-CNN.

Abstract

Wheat protein content is a major determinant of its usage and value. Current methods require wet labs that may be difficult to access and are not real-time. To overcome this, Hyperspectral imaging (HSI) has been reported for estimating the protein content of wheat seeds with the advantage that it is real-time, does not require wet labs, and has high accuracy. However, these models have been developed and validated for a small range of protein content, and without considering cultivation regions. This paper reports the extension of the use of HSI for protein estimation for a wider range of protein content and for wheat cultivated in different regions. Hyperspectral images of 621 wheat samples from five regions in India were acquired in the 900-1700 nm wavelength range. The reference protein content of each sample was determined using the Kjeldahl method, with values ranging from 9.5 to 17.25%. Mean spectra were extracted from the hyperspectral images to develop deep learning and conventional machine learning methods, which were validated through 5-fold cross-validation. The experiments showed that the one-dimensional convolutional neural networks (1D-CNN) performed the best, with the coefficient of determination (R²) of 0.9972, root mean square error (RMSE) of 0.0771, and the ratio of performance to deviation (RPD) of 18.81 for the prediction set. This shows that a 1D-CNN model trained using mean spectra can accurately estimate the wheat protein content. This has the advantage of not requiring a wet lab, and being potentially real-time, which could benefit the farmers, traders, and food industry.

Plant phenotyping relevance

小麦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1D-CNNで推定する手法の拡張・検証が中心であり、交差検証による性能評価も実施している。

Code and data availability

The paper's hyperspectral images, mean spectra, and Kjeldahl protein reference values are not publicly deposited; the Data availability statement says the dataset will be shared only upon reasonable request via email. No public code, model checkpoints, or data URLs are provided.

No evidence-backed public reproduction asset is currently recorded.