| In this paper,the first six leaves of the shoots pick from three kinds of tea trees(Jiukeng,Longjing 43,Juhua Chun)were studied.The physiological structure characteristics of the leaves were observed and compared through scanning electron microscopy and optical microscopy;the distribution characteristics of moisture content and the dynamic characteristics of water loss process were analyzed by statistical methods;the rapid detection of moisture content was realized based on visible-near infrared spectroscopy.Then,hyperspectral imaging technology was adopted to visualize the moisture content and distinguish the front and back sides of tea leaves.Multispectral image and depth image were employed comprehensively to study the feasibility of moisture content detection and height correction.This study can provide theoretical and technical support for rapid nondestructive detection of moisture content in tea production process.The main conclusions are as follows:(1)Statistical method was applied to analyze the distribution characteristics of moisture content and the dynamic characteristics of water loss process.According to the standard method of drying and weighing,the moisture content of leaves was obtained,then the single factor statistical analysis was carried out by ANOVA.The results showed that there were extremely significant differences among different drying time,tea varieties and leaf positions(p<0.01);fixing the drying time factor,tea variety and leaf position factors were analyzed by ANOVA,and the results showed that:at each drying time level,the variety factor,leaf position factor and the interaction of these two factors had extremely significant effects on moisture content of tea leaves(p<0.01);Tukey-kramer method was used for multiple comparisons,the results showed that the average and variance of three levels from variety factor and six levels from leaf position factor were different between any two levels at each drying time level.Calculating the water loss rate of leaves in different drying periods,using MANOVA to carry out single factor multivariate statistical analysis,the results showed that the corresponding d values of different leaf positions for each variety and different varieties for each leaf position were all greater than or equal to 1,it indicates that the water loss rates of different varieties or leaf positions were not identical.Through the statistical analysis of tea moisture content,it is found that the moisture content and water loss rate of tea leaves are affected by drying time,variety and leaf position factors.This part of the research provides reference for the optimization of tea production parameters.(2)Visible-near infrared spectroscopy was adopted to detect the moisture content of tea leaves rapidly and nondestructively.ASD FieldSpec4 spectrometer was employed to collect the front and back spectra(350-2500nm)of each leaf.ANOVA was utilized to analyze the front and back spectra of each variety.The results showed that there were significant differences between the front and back spectra of each variety(p<0.05).Fixing the front and back factor,ANOVA was applied to analyze the spectra of different varieties.The results showed that the spectra of different varieties were very different(p<0.01).Fixing the front and back factor and variety factor,ANOVA was applied to analyze the spectra of different leaf positions.The results showed that there were significant differences between different leaf positions(p<0.05)for the front spectra of Jiukeng and front and back spectra of Longjing 43,but there was no significant difference for other spectra.The moisture content prediction model based on partial least square regression(PLSR)was established by chemometrics,the determination coefficient of prediction set(R_P~2)of the models is 0.968.In order to improve the migration performance of the models,G-L fractional differential processing was applied to the raw spectra.The results show that the fractional differential processing can mine the common information between different spectra,thus improving the transfer performance of the models.Through the study of spectral difference and model migration validation,it provides an analysis idea for the construction of tea moisture content detection model which can be applied to the actual production environment.(3)The hyperspectral imaging technology was used to visualize the moisture content of tea leaves.Two hyperspectral cameras,ImSpector V10E(400-1024 nm)and ImSpector N17E(900-1700 nm),were employed to capture the VNIR and SWIR hyperspectral images of tea leaves.The average spectrum of each leaf was extracted to establish the PLSR models of moisture content.The results showed that the R_P~2 values corresponding to the spectra of VNIR and SWIR were all greater than 0.93.In order to eliminate the possible linear drift and random noise during the hyperspectral acquisition,the pretreatments of the spectra were carried out.The results showed that the optimal pretreatment method for VNIR and SWIR spectra were SNV+MSC.For the sake of reducing the redundant spectral information and model complexity,CARS,SFLA and GA were used to screen initial feature bands,and then SPA was used to refine the initial feature bands.The results show that for VNIR front and back spectra,the models based on the feature bands selected by SFLA+SPA is the best,and the corresponding R_P~2 values of LSSVR models are 0.957 and 0.960,respectively.For SWIR front and back spectra,the models based on the feature bands selected by GA+SPA is the best,and the corresponding R_P~2 values of LSSVR models are both 0.995.In order to distinguish the tea leaves lying on the belt are facing upward or backward,classifiers based on bands ratio were established.The results show that the classification accuracy of VNIR spectra based on LDA classifier is 99.4%,for SWIR spectra,the accuracy is 96.7%.In order to realize the imaging of moisture content distribution for overlapped tea leaves,the algorithm based on spectral angle statistics was adopted to segment the hyperspectral image,calculating the average spectrum of each region,identifying the side of each region according to the average spectrum,then importing the feature bands of each pixel into the corresponding model,and the visualization of tea leaves moisture content distribution was realized finally.This part of the research provides a method support for realizing the visual detection of tea moisture content.(4)The feasibility of moisture content detection and height correction was studied by using multispectral image and depth image.Ximea multi-spectral camera and Kinect 2.0 camera were used to collect the multispectral image and depth image of the tea leaves respectively,then vignetting correction and black-and-white board correction were carried out for the multispectral image at each band.After extracting the spectra from the corrected multi-spectral image,PLSR,LSSVR and ELM models were established respectively to predict moisture content.The results show that the LSSVR model is optimal,for the front and back spectra,the R_P~2 values are 0.771and 0.681 respectively.Collecting the white board spectral information at different detection heights,and the height correction parameters of different bands were calculated according to the empirical formula of spectral attenuation.To obtain the detection height information automatically,Zhang Zhengyou's calibration method was adopted to correct the distortion of multispectral image first,regarding the corrected multispectral image as the reference registration image,and the depth image as the pending registration image,the registration operation based on SIFT feature points was carried out.Finally,the registration of multispectral image and depth image was successfully realized.According to the this scheme,the spectral correction of tea leaves can effectively improve the accuracy of moisture content prediction,which provides a train of thought for the detection height correction of multispectral image which was captured through area-scanning. |
PDF Full Text Request