Establishment And Application Research Of Fresh Tea Hyperspectral Database

In tradition,the identifying tea species is by leaf feature which needs some experience.Tea quality is evaluated by doing some kinds of chemical experiment which is trivial operation,time-consuming and laborious and also involves high cost.Hyperspectral technology can detect the subtle difference between ground objects and contain vast amount of information.It is meaningful to use hyperspectral technology in tea recognition and quality evaluation.Building in-situ tea hyperspectral database and implementing data storing,management and sharing are the basis of tea spectral study.In this thesis,tea canopy and leaf spectral were tested and collected and the database application system is developed using python program language.The library,such as PyQt5.6,numpy,pandas etc are used in this system,and PostgreSQL 9.5 is the background database.Data were selected and processed using the built system.The tea spectral features were investigated by SVM-RFE and Elastic Net algorithm.Two classifiers were trained using those features for tea species identify.The main work and achievements are as follows:The business process,modules,architecture,data specification,database structure of the system was designed based on target users' needs.Based on C/S architecture,using PostgreSQL 9.1 as background database and python language and some python libraries,such as PyQt 5.6,numpy,pandas etc,build the tea hyperspectral database application system.The system includes some import functions,such as ASD FiledSpec3 binary data reading and writing,data management and display.The System also integrated many useful functions,such as SG data smoothing,spectral derivation,spectral inverse log conversion,spectral parameters and spectral indices calculation etc.To avoid taking time in picking up "good" data,proposal an automating picking up line algorithm based on experiment threshold which obviously reduce the data preprocessing time costing.Local data sets could be seamlessly used with data from database.Data in the list could be exported as a CSV file whose band ranges were assigned by user.Select 9 kinds of tea hyperspectral data from the database which were detected in 8,Oct,2014.The integrated functions were used to preprocess,the removal of water absorb bands for instance,and calculate 24 spectral indices.SVM-RFE and Elastic Net(EN)were used as feature selection method.Linear support vector machine(SVM)and random forest(RF)were used to separate the species by using preprocessed reflectance data,spectral indices and feature selected dataset of both,respectively.The results of this study suggest that SVM-RFE and EN is stable and efficient feature selection algorithm and the Linear-SVM classifier outperforms RF.Some of the Features selected by SVM-RFE and RF located in the same region,but others not.In the SVM classifier,the overall accuracy was greater than 90%.In the RF classifier,the accuracy in the spectral dataset just was 70%,but in the indices dataset the accuracy was over 90%.