Non-destructive Evaluation Of Blueberry Quality Using Hyperspectral Imaging

Optical non-destructive techniques have been considered as promising scanning measurements for the detection of horticultural and food products.Small berries including blueberry and strawberry are extensively consumed fruits with great economic values due to their characteristic flavor and appearance as well as potential health benefits.Due to the perishable characteristics in nature,small berries have been and still being suffered enormous economic loss.Use of the optical non-destructive techniques for the control of the quality and safety of small berry fruits can provide the consistent and safe as well as nutritious products for consumers,thus greatly increasing their economic values.Hence,this dissertation attempt to explores methods for the non-invasive evaluation of blueberry quality by the application of spatial and spectral data respectively obtained from hyperspectral imaging and NIR spectrograph.The optical non-destructive techniques and their applications for quality and safety control of small berry fruits were first elaborated,including Vis-NIR spectroscopy,computer vision system,hyperspectral imaging,multispectral imaging,laser-induced method and thermal imaging.The discussion with respect to the photoacoustic technique,X-ray technique,Terahertz spectroscopy,odor imaging,micro-destructive testing and smart mobile terminal-based analyzer,and their potential applications in agricultural and food domains were also presented.Among these techniques,the hyperspectral imaging can obtain both spatial and spectral information,and therefore,it allows the comprehensive evaluation of the testing samples.Hence,a custom-made hyperspectral imaging system containing reflectance and transmittance modes for simultaneously estimating several mechanical properties of blueberry non-destructively.The reflectance and transmittance wavelengths were extracted from the corresponding segmented hypercubes,and were related to the mechanical properties of blueberries calculated from texture profile analysis and puncture analysis via the use of least squares-support vector machine(LS-SVM).A state-of-art variable selection approach termed random frog was applied to gather wavelength specific to mechanical properties.Prediction models using random frog selected reflectance and transmittance wavelengths yielded the comparable results to these using the respective entire wavelengths.The combined wavelengths combining reflectance and transmittance with single random frog had the feasibility for the estimation of predicting hardness,springiness,resilience,force max and final force,with Rp(RPD)values of 0.86(1.78),0.72(1.73),0.79(1.78),0.77(1.51)and 0.84(1.72),respectively.The combined wavelengths with double random frog could give satisfactory prediction models with fewer spectra.In conclusion,hyperspectral reflectance and transmittance in tandem with random frog variable selection demonstrated a considerable potential for the prediction of blueberry mechanical properties simultaneously.For the purpose of validating the feasibility of interacted spectra for non-destructively estimating comprehensive blueberry mechanical properties.A suspended module was added to the hyperspectral reflectance and transmittance system,thus allowing the acquisition of blueberry interactance hypercubes.A region growing based algorithm was used to segment the obtained interactance hypercubes and to assist in the extraction of mean interacted spectra.Afterwards,the interacted wavelengths were pre-processed by Standard Normal Variate(SNV)and Savitzky-Golay first derivative(Der),and LS-SVM associated with Monte Carlo-uninformative variable elimination(MC-UVE)was applied to build the prediction models using smoothed spectra for blueberry mechanical parameters.Based on the MC-UVE selected wavelengths,the SNV model performed best for cohesiveness with Rp(Rc)value of 0.91(0.91).The SNV models of springiness,resilience,max force strain and final force resulted in Rp(Rc)values of 0.84(0.85),0.86(0.87),0.65(0.76)and 0.62(0.72),respectively.The Rp(Rc)values of Der models were found to be 0.77(0.86),0.71(0.73)and 0.58(0.69)for hardness,maximum force and gradient,respectively.The overall performances of MC-UVE based models were in general similar to those with full spectra.The above results showed the potential of hyperspectral interactance imaging coupled with MC-UVE approach for predicting the blueberry mechanical properties.However,the most of publications established the prediction models using the certain cultivar harvested from the certain season,not taking the biological variations into account.To evaluate the effects of biological variations on the robustness of established prediction models,three scanning tools viz.hyperspectral reflectance and transmittance sensing as well as near infrared(NIR)spectroscopy were applied to predict blueberry postharvest quality such as firmness,elastic modulus and soluble solid content(SSC)regarding biological variability inclusive of cultivar and season.LS-SVM models were established from these three spectra based on samples from the three cultivars viz.Bluecrop,Duke and M2 and two harvest years viz.2014 and 2015.One-cultivar reflectance models(establish model using one cultivar)derived better results than the corresponding transmittance and NIR models for predicting blueberry firmness with few cultivar effects.Two-cultivar NIR models(establish model using two cultivars)proved to be appropriate for the estimation of blueberry SSC with correlations over 0.83.Rp(RMSEp)values of the three-cultivar reflectance models(establish model using 75% of three cultivars)were 0.73(0.094)and 0.73(0.186)for predicting blueberry firmness and elastic modulus,respectively.For SSC prediction,the three-cultivar NIR model produced the Rp(RMSEp)value of 0.85(0.090).Adding Bluecrop samples harvested in 2014 could enhance the robustness of three-cultivar model of firmness and elastic modulus.However,the performance of some prediction models was likely to decline via the addition of more samples.The above results indicated that the possibility of using spatial and spectral techniques to develop the robust models for predicting blueberry postharvest quality containing biological variability.The in-depth studies are required for compromising the robustness and accuracy of model.During the harvest,transportation and storage,blueberry will produce the non-visible mechanical damage which cannot be detected by the human naked eye and traditional computer vision system.In addition,with the application of hyperspectral imaging technique,the most of current publications inspected the mechanical damages not considering the time factor.Thus,the hyperspectral reflectance and transmittance as well as interactance data in tandem with chemometrics methods were used to detect this non-visible blueberries mechanical damage with time evolution.The hypercubes were first automatically segmented by the region growing based algorithms,and the maximum-normalized spectra were pretreated by the Standard Normal Variate algorithm.Subsequently,Competitive Adaptive Reweighted Sampling algorithm was applied to extract the damage-specific wavelengths.According to the confusion matrices and area under Receiver Operating Characteristics(ROC)curves,transmittance demonstrated relatively superior performance to reflectance and interactance.Application of new blueberry sample set which was subjected to impact tests with the time after damage,results showed that it was especially difficult to distinguish fresh damage in blueberry.For 2 days after impacted,several transmittance-based classifiers obtained satisfactory accuracies for classifying damaged(sound)blueberries – logistic regression 79.1%(85.7%),multilayer perceptron-back propagation 74.4%(92.1%)and logistic function tree 72.1%(95.2%).In addition,the physical property preliminarily proved to be more pronounced than the absorbed impact energy for damage incidence and severity of blueberry through the use of multiple comparison.This dissertation has demonstrated the usefulness of spatial and spectral based techniques viz.hyperspectral and NIR spectroscopy for the assessment of blueberry quality.Further researches are needed to exploit the feasibility of the other optical based methods inclusive of the photoacoustic technique,odor visualization and micro-destructive testing for blueberry quality and safety control,and apply these techniques to the other agricultural and food products as well as biological materials.Furthermore,we have to focus on the achievement of these techniques for real-time,on-line and point-of-sale detection of fruit quality and safety both for industrial and fundamental applications.