Study on Near Infrared Spectroscopy Detection Model for Oil Content of Cyperus esculentus Tubers

Wei Haifeng; Shi Xueshuang; Dang Xiqiang; Mimadunzhu; Zhang Mengyuan; Chang Wei; Tinley Tsomo; Zhang Bin; Gao Wenwei

doi:10.14083/j.issn.1001-4942.2025.01.023

Abstract：In order to establish a rapid non-destructive detection model of near-infrared spectroscopy for oil content of Cyperus esculentus tubers and to improve the efficiency of early generation selection of breeding materials,109 samples of C.esculentus tubers were used as experimental materials in this study.The near-in-frared spectra with a wavelength range from 950 to 1 650 nm and a resolution of 1 nm were collected,and the crude fat content of the tubers was determined by Soxhlet extraction method.After eliminating abnormal sam-ples,a total of 103 samples were obtained which were then divided into calibration set and validation set in a ratio of 3∶1 by SPXY method.The original spectra were preprocessed by standard normal transformation,mul-tiple scattering correction,first derivative,second derivative,SG smoothing and their hybrid methods,respec-tively,to establish the partial least squares regression(PLSR)model.Through comparing and analyzing the performance of the models,the MSC+SG method with better pretreatment effect on both the calibration set and the validation set was selected for the construction of the oil content detection model of C.esculentus.Then,the competitive adaptive reweighted sampling(CARS),uninformative variable elimination(UVE)algorithm and MLP neural network were used to extract the characteristic wavelengths,and the PLSR model was constructed.The results showed that 115 and 251 characteristic bands were extracted by CARS and UVE algorithms,re-spectively,with the modeling effect better than that of the full-band.The CARS-PLSR model had the best pre-diction performance with the root mean square error(RMSE)of cross-validation(RMSECV)and coefficient of determination(R2CV)of the calibration set were 1.328 and 0.903 respectively,and the RMSEP and RP of the validation set were 1.206 and 0.888 respectively,and the relative analysis error(RPDP)of the validation set was 3.040.The prediction accuracy of the MLP-PLSR model was close to that of the CARS-PLSR model with RMSECV and R2CV were 1.387 and 0.903,RMSEP and RP were 1.207 and 0.887,and RPDP was 3.040,but the extracted characteristic wavelengths were only 77 which was the fewest among the three methods,indicating that the MLP method could more effectively reduce the spectral information overlap and filter out irrelevant in-formation,and was more suitable for the oil content detection of C.esculentus.In summary,this study prelimi-narily established a rapid and non-destructive detection model of oil content in C.esculentus tubers based on near-infrared spectroscopy,which could provide an effective method for improving the detection efficiency for breeding work and provide technical support for non-destructive detection of oil content in C.esculentus.