Research On The Detection Method Of Rice Seed Respiration And Its Vigor Grading Model

Rice is one of the vital crops in China,high-quality seeds are the prerequisite to maintain sustainable agricultural development.Seed vigor is used as indicator to evaluate seed quality,however,there are defects in traditional seed vigor detection,such as long detection time and damage to the seeds.Respiration,as a important physiological process of plant seeds,is closely related to seed vigor.Tunable Diode Laser Absorption Spectroscopy(TDLAS)can be used for real-time measurement of carbon dioxide concentration produced by rice seed respiration,for it is a trace gas detection method with high sensitivity.In this study,a nondestructive rice seed vigor grading method based on TDLAS and deep learning was proposed,and a seed vigor grading model which acted directly on the time-domain data of seed respiration was designed,rice seeds VLiangyou1219 and Yong You9 were taken as samples.The main research is as follows.(1)The influences of modulation frequency,modulation amplitude,scanning frequency and scanning amplitude on the amplitude of second harmonic signal were studied,the optimal parameters were used in TDLAS system through simulation.Rice seeds after accelerated aging method at high temperature and high humidity were exploited,and the vigor indexes of the seeds were measured by the standard germination test.The eight hours' seeds respiration data during germination were collected by the TDLAS system,and the collected data were preprocessed to establish the seed respiration datasets.(2)Transformer does well in feature extraction of time series data,though the high requirements for memory and computing resources limit its application,in this thesis,an end-to-end1 D convolution Transformer-based seed vigor grading model was studied.The results revealed that the proposed model achieved 90.74% and 95.83% classification accuracies on the test set of two hours respiration data of VLiangyou1219 and Yongyou9,respectively,which showed an improvement compared to the Transformer model,and outperformed the other models including the self-built fully connected neural network,deep learning and various machine learning models.On the test sets of four hours later respiration data of VLiangyou1219,the accuracies of the proposed model were higher than the others,on the test sets of four hours later respiration data of Yongyou9,the accuracies of different models were similar.The classification results of the proposed model were visualized by t-SNE,showed that the seed respiration data features were distinguished significantly by the last fully connected layer.(3)According to the researches,a seed vigor grading software based on seed respiration data was developed using MATLAB and Python with Tkinter framework.The software consisted of TDLAS system simulation module,seed respiration data preprocessing module,seed respiration spectrum analysis module,seed vigor grading module,log saving module and help documentation.The seed vigor grading method based on seed respiration could achieve the goal of seed vigor grading effectively and quickly,and the seed vigor grading software could promote the efficiency in practical applications.