Study On Seed Quality Detection Based On Hyperspectral Technique

As the main plant and crop production,seed has great biological and economic significance,and has been highly concerned by farmers,producers,seed management enterprises,seed quality supervision and inspection stations and seed management departments.Seed quality is closely related to people’s health and living standards,and determines the level of agricultural development.Traditional methods are difficult to meet the requirements of batch seed quality detection in modern agriculture and forestry.Seed variety,seed vigor and seed moisture content are the main quality indexes of seeds.This paper studies the seed quality detection methods from three aspects: variety,vigor and moisture content based on hyperspectral technology taking corn and soybean seeds as experimental objects.The main research contents are as follows:First,a hyperspectral database of seeds of different varieties,different vigour and different water content was established.In the variety identification experiments,samples were selected from 10 maize seeds and five soybean seeds that were similar in appearance and difficult to identify with the naked eye;in the study of seed vigour and seed moisture content,samples were still selected from maize and soybean seeds.In the seed vigour test,the two seeds were artificially aged using different temperatures and times,and the seed vigour levels were verified by standard germination tests.In the seed water content experiment,a 10 gradient water absorption experiment was designed.After obtaining the samples,a hyperspectral image acquisition system was constructed and the spectral data(370-1042 nm)of all seeds were collected to establish the spectral database and image information of the samples,which provided data support for the subsequent theoretical analysis and algorithm research.Secondly,a method for seed variety identification based on integrated Bagging learning is proposed using hyperspectral techniques.The spectral data of different varieties of seeds were studied to compare and analyse the reasons for the differences in spectral reflectance curves.After obtaining the sample spectral data,data correction was carried out,followed by MSC,SG and SNV preprocessing respectively,and the feature bands were extracted using PCA and SPA.Finally,the variety recognition models of maize and soybean seeds were established based on ELM,KNN,SVM,DT and Bagging respectively,and the recognition results of the models were compared under different preprocessing and feature dimensionality reduction.Based on the full-spectrum + MSC + SVM model,the accuracy was 99.44% and 98.33% for the training and test sets of maize seeds,respectively,and94.38% and 89.11% for soybean seeds,respectively.the accuracy of the DT and Bagging methods was above 98.75% for both the training and test sets of the two types of seeds.Then,using hyperspectral techniques,a method of seed vigour detection based on Bagging integrated learning is proposed.By studying the spectral data of seeds with different vigour classes,the reasons for the differences in spectral reflectance curves were compared and analysed.After obtaining the sample spectral data,data correction was carried out,followed by data pre-processing and feature downscaling.Finally,seed vigour detection models were developed based on ELM,KNN,SVM,DT and Bagging,and the detection results of the models were compared under different preprocessing and downscaling methods.The full-spectrum SVM model achieved 100% and 98.61%accuracy in the training and test sets for maize seed vigour detection,respectively,and 100% accuracy for soybean seeds,both of which were better than the ELM and KNN models.Finally,a method for detecting the water content of seeds based on least squares discriminant analysis was implemented using hyperspectral techniques.The spectral data of seeds with different water contents were studied and the reasons for the differences in spectral reflectance curves were compared and analysed.After obtaining the sample spectral data,data correction was carried out,and then data pre-processing was performed on the original spectra,followed by feature wavelength extraction using the correlation coefficient method,the UVE method and the CARS method,and a partial least squares detection model was established,and the detection results of the model were compared under different pre-processing and feature downscaling.For maize seeds,the SNV-based partial least squares model had the best accuracy and error.The correlation coefficients were 0.880 and 0.855 for the calibration and validation sets respectively,with a root mean square error of 2.18%for the calibration set and 2.44% for the validation set.the CARS method was able to achieve maximum convergence of errors.The validation set correlation coefficient for the moisture detection model based on 25 wavelengths was 0.915 and the validation set RMS error was 2.29%.For soybean seeds,SG smoothing combined with first order differential pre-processing of the spectra gave the best robustness.The calibration set correlation coefficient was 0.946 and the validation set correlation coefficient was 0.941.The root mean square error of the calibration set and validation set were 1.435%and 1.473%,respectively.27 characteristic wavelengths were screened by the CARS method to develop a quantitative soybean moisture content detection model,and a calibration set correlation error of 0.9719,a calibration set root mean square error of 1.045% and a validation set The correlation coefficient was 0.9621,and the root mean square error of the validation set was 1.174%.The method and the conclusions obtained in this study are of great significance for rapid nondestructive testing of seed quality,and can also provide reference for rapid non-destructive testing of seed quality of other crops.