Peanuts Variety Classification,Mildew Detection And Protein Content Predication Based On Hyperspectral Image

Peanut is an important edible oil and non-staple food processing raw material crop in daily life.In 2021,the global output of peanut reached 50.63 million tons,and its economic value is significant.It is widely grown in more than 100 countries and regions in the world.There are more than 500 kinds of peanut varieties in China,and more than 30 kinds of excellent varieties,providing high-quality seed source for peanut cultivation in the world.For classification of peanut varieties,mildew detection and protein content detection,the traditional chemical measurement have certain loss,time-consuming,laborious and other significant problems,therefore,this research carried out hyperspectral imaging technology to achieve non-destructive and rapid detection of peanut quality,for which provides a new theory and method for peanut quality identification.The main results are as follows:(1)A peanut variety classification model based on hyperspectral image was constructed.Spectral images of five peanut varieties were collected.According to the strong correlations between the spectral data bands and the amount of total high redundancy feature,spectral data preprocessing is realized by a variety of algorithms.e Xtreme Gradient Boosting(XGBoost),Light Gradient Boosting Machine(Light GBM),Categorical Features Gradient Boosting(Cat Boost)and Gradient Boosting Decison Tree(GBDT)algorithms were used to extract feature bands,and then XGBoost and Light GBM were used to conduct classification model for the extracted feature bands.In this model,both XGBoost and Light GBM can achieve 99.33 % and the other performance of the two algorithms were very similar,so the Optuna algorithm was selected to optimize the two algorithms respectively.After optimization,the performance indexes of Light GBM and XGBoost algorithms were improved in different degrees,and Light GBM was relatively obvious.In particular,fit time index was improved huge compared with XGBoost algorithm.Light GBM algorithm revealed that 401 nm was the best band for classification.The best classification model was MF-Light GBM-Light GBM-Optuna-Light GBM.(2)A mildew peanut detection model based on hyperspectral image was constructed.120 peanuts from each of the 5 kinds of peanuts purchased in the market were selected,a total of 600 peanuts were selected for mildew treatment.400 peanuts from each kind were randomly selected,a total of 2000 peanuts were sterile preserved.After 30 days,the spectral images of all peanut samples were collected.Then variety of pre-processing algorithms were used to reduce the noise,and Median Filtering(MF)was the best,and its recognition rate have reached 97.74 %.Using GBDT,Light GBM,Cat Boost and XGBoost algorithms to extract important feature bands in the spectral data preprocessed by MF,it can be seen that the best recognition band was near 1000 nm.Using GBDT,Light GBM,Cat Boost and XGBoost algorithms to model the extracted the feature bands preprocessed by MF.The results showed that Light GBM was the best algorithm,and its detection rate was 99.10%.Optuna algorithm was used to tune its parameters,which reduced the time of model.The samples were divided into mildew and healthy peanuts,and the recognition result reached100 %,the running time was only 0.084 s.Therefore,the optimal detection model MFLight GBM-Light GBM-Optuna-Light GBM was obtained.(3)The prediction model of peanut protein content based on hyperspectral image was built to collect the samples after using spectrophotometry and measured its protein content.Spectral data preprocessing is realized by a variety of algorithms.XGBoost,Light GBM,Cat Boost and Ridge algorithms were used for feature extraction and modeling of peanut spectral data to obtain the optimal model,MF was used for spectral data preprocessing,XGBoost was used for feature band extraction.Ridge was used to model the extracted characteristic bands and measured protein content.And the hyperparameter α of the Ridge algorithm and XGBoost parameters were tuned to obtain the best model MF-XGBoostRidge-hyperparameter α tuning-Optuna-XGBoost-Ridge.The evaluation indexes Mae=0.007,Mse=0.00001,Med AE=0.0065,Rmse=0.009,correlation coefficient=0.886 and fit time was only 0.02 s,which greatly saved the calculation time compared with the traditional model.