Building hyperspectral model of oleic acid content in rapeseed of vegetable and oil type Brassica napus

【Objective】Based on the hyperspectral characteristics of Brassica napus, a preliminary discrimination of rape moss picking was conducted, and provided theoretical guidance for the realization of hyperspectral inversion of rapeseed oleic acid content.【Method】The spectral data of B. napus leaves were collected by FieldSpec 3, and the oleic acid content of rapeseed with mossed treatment and unmossed treatment was analyzed by Agilent GC-MS 7980 B gas chromatograph. The average original spectral reflectance characteristics of the two treatments were compared. The correlation between the original and first-order differential spectral reflectance and the oleic acid content of rapeseed with and without moss picking was analyzed. On the basis of the original spectrum characteristic wavelength based support vector machine(SVM) discriminant model, based on spectral parameters of the oleic acid content in binomial model, based on the firstorder spectrum characteristic wavelengths of oleic acid content(MLSR) with multiple linear stepwise regression and partial least-squares regression(PLSR) prediction model, independent samples T test was used to verify this model precision.【Result】It was found that the average original spectral reflectance curves of unpicked and moss-picked rapeseed were different in the band range of 760-1080 nm. There were some differences between the original spectral reflectance and the oleic acid content of rapeseed between unplucked and plucked rapeseed. The original spectral reflectance of unplucked rapeseed was 484-956 nm and 1001-1146 nm, and the original spectral reflectance of plucked rapeseed was 1882-2111 nm and 2324-2499 nm, indicating that the correlation between spectral reflectance and oleic acid content of rapeseed was affected by moss picking. The original spectral reflectance at four inflection points(760, 920, 970 and 1080 nm) in the 760-1080 nm band was selected as the independent variable to construct the SVM discriminant model. After multiple random sampling and comparison of all SVM discriminant models, it was found that the overall accuracy of the training set sample of the best discriminant model was 86.1%. The overall accuracy of the validation set was 77.8%, which indicated that it was feasible to use hyperspectral technology to determine whether rapeseed moss picking. Among the spectral parameter models, RVI model had the best inversion effect on the content of rapeseed oleic acid in unmossed rapeseed, and the correlation coefficient between RVI and the content of rapeseed oleic acid in unmossed rapeseed was-0.705. Comparing the prediction models of oleic acid content in rapeseed of all types, PLSR model had the highest prediction accuracy of oleic acid content in rapeseed of unmossed rapeseed, with its training set R²=0.590 and RMSE=0.610;MLSR model had the highest prediction accuracy of oleic acid content in rapeseed of mossed rapeseed, with its training set R²=0.773 and RMSE=0.874. The independent sample T test was used to verify the two model test sets of samples. P=0.839 for the unmossed sample and P=0.858 for the mossed sample, and there was no significant difference between the measured and predicted values of the two samples(P>0.05). The model was reasonable, indicating that it was feasible to predict the oleic acid content of rapeseed using hyperspectral technology.【Suggestion】Introduce random forest and other machine learning algorithms to better select characteristic wavelengths(significantly correlated wavelengths or full bands, etc.) and improve the prediction ability of spectral data for rapeseed oleic acid content. Later experiments should focus on the estimation of rapeseed oleic acid content in multiple varieties of B. napus, and explore whether hyperspectral technology is universally feasible to estimate rapeseed oleic acid content. Other rapeseed quality indexes were retrieved by hyperspectral technique, which provides the basis for monitoring rapeseed quality by hyperspectral remote sensing.