Research And Application Evaluation Of Corn Stalk NDF Content Prediction Model Based On Machine Learning

Neutral detergent fiber(NDF)content is the critical indicator of fiber in corn stover,and its detection plays an important role in the quality analysis of corn stover.In view of the problems existing in the detection of NDF content of corn stover found by predecessors based on near infrared spectroscopy,such as the lack of common technique of spectral analysis,the lower accuracy of prediction model and so on.This study carried out spectra handling and developed a prediction model to precisely measure NDF content using kinds of spectral analysis methods and machine learning algorithms.The specific research contents and results are as follows:(1)The spectral data of 530 corn stover samples and corresponding NDF content values were obtained in this research.Firstly,outliers were picked out using the Mahalanobis Distance method and Monte Carlo Cross Validation(MCCV)algorithm,the effective outliers were deleted to ensure the correctness of the original data after their characteristic was investigated by checking one by one.Then,the stratified sampling method was employed to divide the sample sets to improve the representativeness of the samples,training set and test set were formed.Next,derivative method,multiple scattering correction(MSC),Savitzky Golay smoothing(SG smoothing)and their combination methods were separately utilized to prepossess the spectra,and the effect of pretreatments were compared.(2)Furthermore,for reducing the number of redundant variables,Correlation Coefficient method(CCM),Synergy interval Partial Least Square(SiPLS),Genetic Algorithm(GA)and cascaded algorithm CCM-GA and SiPLS-GA were discussed and compared their effects on extraction of feature wavelength.Finally,the predictive models using partial least square regression(PLSR)and support vector regression(SVR)were established to detect NDF content and compared.The optimized number of principal component for PLSR model was determined by global search algorithm,regularization parameter(C),insensitive parameter(?) and kernel parameter of RBF(?) for SVR were optimized by using grid search.The results showed that PLSR model had higher predictive accuracy for detecting NDF content,the evaluation parameters were calculated with R~2?RMSEP and RPD attaining 0.9756,0.4586%and 6.45,respectively.In addition,20 unknown samples were utilized to test the generalization of PLSR model with the mean of predictive bias obtaining 0.2425%.The research results demonstrated that the established prediction model for NDF content detection using kinds of spectral analysis methods and machine learning algorithms,had higher prediction accuracy to accomplish the measurement of NDF content.The proposed design scheme provides a theoretical reference for the automation of detecting NDF content,which has positive significance for analysis of corn stover quality.