Research On Forest Seed Vigor Detection And Prediction Method Based On Hyperspectral And Deep Learning

Seed vigor is an important index of seed quality,which has a great influence on the speed of seed germination and seedling health.Non-destructive testing of seed vigor plays an important role in optimizing the cost of crop planting.Therefore,the establishment of a fast,non-destructive,and highprecision method for detecting seed vigor has important biological and economic significance for ensuring seed quality and increasing crop yields.In this paper,based on the hyperspectral imaging technology,two forest seeds(Quercus variabilis and Sophora japonica)were taken as the experimental objects to carry out the research on the method of rapid detection and prediction of seed vigor.The proposed autoregressive local optimization-successive projections algorithm combined with machine learning and deep learning algorithms to realize the identification of seed vigor,and the correlation between spectrum and components of different vigor seeds was established.The main research contents and conclusions of this article are as follows:(1)Designed and realized the construction of a hyperspectral image acquisition system,and established a hyperspectral database of different vigorous seeds.According to the characteristics of the experimental object,a complete hyperspectral imaging system including hyperspectral imager,mobile platform and host computer interface was built to collect the data of Quercus variabilis seeds and Sophora japonica seeds with high vigor,low vigor,and non-viable samples.This step continued in the first 10 hours of the germination process.The collected spectra and image information are used to establish a hyperspectral database of the samples to provide data support for subsequent theoretical analysis and algorithm research.(2)The rapid detection of seed vigor was realized and an autoregressive local optimizationsuccessive projection algorithm was proposed.Analyzed and studied the spectral response of seeds with different vigor.Four typical machine learning algorithms were used to achieve spectral-based seed vigor detection,and the effects of different spectral preprocessing methods and feature selection methods on model performance were compared.The results show that the accuracy of the model built with the original spectral data set is 70%-80%.The use of appropriate preprocessing methods can effectively improve the accuracy,up to 99.94%,and the selection of characteristic bands can eliminate redundant information to a certain extent to improve model performance.(3)According to the spectral and image characteristics of different vitality seeds,it is proposed to use one-dimensional and two-dimensional deep learning networks to realize seed vitality detection,and introduce particle swarm optimization algorithm to optimize the model.1DCNN,2DCNN and LSTM were constructed for seed vigor detection.Particle swarm optimization was used to optimize the learning rate and batch size to improve the accuracy of the model.The results show that the ability of deep learning structure to detect seed vigor is better than machine learning algorithm.And the image-based model has the best effect(recognition rate is close to 100%).However,feature pre-extraction and selection will significantly reduce the recognition ability of deep learning models.(4)Proposed and realized the early and rapid prediction method of seed vigor based on hyperspectral.Seed vigor detection was achieved by collecting complete hyperspectral data within 10 hours of germination of different vigor seeds.Based on the structure of the curve of the average spectra of seeds with different vigor with the germination time,it can be found that Quercus variabilis seeds and Sophora japonica seeds with different vigor are clearly distinguished at 0h of germination.At the same time,using the optimal vigor detection model to analyze the cumulative contribution rate of the model at each germination time point,it was found that the accuracy rate decreased with germination,and the model performed better at 0h germination and required less data.Therefore,the data of 0h of seed germination can be used to realize the early prediction of seed vigor.(5)Through research and qualitative analysis of the main components and average spectra of seeds in different aging and germination states,the correlation between seed vigor,spectrum and internal main components was established.Artificial aging and water-absorbing germination could change the content of main components(water,protein,starch,fat,and total sugar)inside the seed.The significance analysis of each component between adjacent aging treatments was established,and Pearson correlation coefficient was used to establish the correlation between spectra and components.In summary,it is found the protein in the seeds of Quercus variabilis with different vigor is the most sensitive to the 850-1000 nm spectral changes,while the changes in seed vigor of Sophora japonica have a greater impact on the changes in seed fat content and spectrum.The methods and results of this study has a momentous role on the rapid identification and early prediction of seed vigor,and have reference value for improving the non-destructive detection ability of hyperspectral imaging technology for different species and different vigor seeds.