Using VIS-NIR hyperspectral imaging and machine learning for non-destructive detection of nutrient contents in sorghum

Abstract

Abstract Non-destructive, rapid, and accurate detection of the nutritional compositions in sorghum is of great significance to the application of sorghum in agricultural production and food industry. In the process of sorghum nutrition detection, it can obtain good effect by extracting the corresponding characteristic wavelengths and selecting the suitable detection model for different nutrients. In this study, the crude protein, tannin, and crude fat contents of sorghum variety samples were taken as the research object. Firstly, the visible near-infrared(Vis-NIR) hyperspectral curves of sorghum were measured by the Starter Kit indoor mobile scanning platform (Starter Kit, Headwall Photonics, USA). Secondly, the nutritional components were determined using chemical methods in order to analyze the differences in nutritional composition among different varieties. Thirdly, the original spectral curves were de-noised by Standard normal variate(SNV), Detrending, and Multiplicative Scatter Correction (MSC) algorithms, and the Competitive adaptive reweighted sampling (CARS) and Bootstrapping soft shrinkage (BOSS) algorithms were used to coarse extract the characteristic variables, then Iteratively retains informative variables (IRIV) was used to judge the importance of the characteristic variables, and the optimal wavelength sets of crude protein, tannin and crude fat were obtained respectively. Finally, Partial least squares(PLS), Back propagation(BP) and Extreme learning machine(ELM) were used to establish the non-destructive detection models of crude protein, tannin and crude fat content respectively. The results showed the following: (1) The optimal variable sets of crude protein, tannin and crude fat contain 41, 38 and 22 wavelength variables, respectively. (2) The CARS-IRIV-PLS model was suitable for detecting crude protein, the prediction set exhibits R 2 , RMSE and RPD values of 0.6913, 0.7996% and 1.7998. The BOSS-IRIV-PLS model achieved good results in tannin detection, the prediction set exhibits R 2 , RMSE and RPD values of 0.8760, 0.2169% and 2.8398. The BOSS-IRIV-ELM model achieved the best results in crude fat detection, the prediction set exhibits R 2 , RMSE and RPD values of 0.6145, 0.3208% and 1.6106. (3) Linear PLS model is suitable for crude protein and tannin detection, and nonlinear ELM model is suitable for crude fat detection. These detection models can be used for the effective estimation of the nutritional compositions in sorghum with Vis-NIR spectral data, and can provide an important basis for the application of food nutrition assessment.

Plant phenotyping relevance

ソルガムの栄養成分含量をVIS-NIRハイパースペクトル画像と機械学習で非破壊推定する測定・解析手法が研究の中心であり、複数モデルの性能評価も行っている。

Code and data availability

The paper describes VIS-NIR hyperspectral imaging of 279 sorghum samples and machine-learning models, but provides no public dataset, image, code, or model deposit. The Data Availability statement only claims data are included in the article itself, with no repository, URL, or accession. All listed URLs are cited prior

No evidence-backed public reproduction asset is currently recorded.