Study On Non-destructive Detection Of Peach Quality Parameters Based On Hyperspectral Fluorescence Imaging Technology

With high economic value,good medicinal value and a variety of nutritional values required by the human body,peach has always been popular among consumers of all ages.Sugar,firmness and p H are important parameters related to the quality of peach,which are the main factors determining the freshness of the fruit and directly affect consumers’willingness to buy.Although the results are accurate,the process is tedious and can cause damage to the peach fruit,so an efficient,non-destructive and accurate method is needed to test the quality of peach.Therefore,this paper selects hyperspectral fluorescence imaging technology to examine the quality of peach,and investigates the feasibility of non-destructive detection of various internal quality parameters of peach.(1)For the sugar detection of peach,the spectral data were collected at the top,near the equator and at the bottom of peach using a hyperspectral fluorescence imager.The PLSR and ELM models were combined to model the different acquisition locations,and the best performance of the raw data near the equator was found after comprehensive analysis and comparison based on the prediction index.Six pre-processing methods such as Detrending(DT)and Gaussian filtering(GF)are applied to the raw spectral data near the equator,and the experimental results show that the GF method has the best noise reduction effect.Five primary feature extraction methods such as Bootstrap Soft Shrinkage(BOSS)and three secondary feature extraction methods such as VISSA-MASS were selected to screen the feature variables.Three models,random forest regression(RFR),particle swarm optimization-based least squares support vector machine(PSO-LSSVM)and partial least squares regression(PLSR),were developed for the feature variables extracted by the eight different methods.The results show that the GF-VISSA-MASS-PLSR method based on the equator has the best prediction effect,with R_p~2,R_c~2and RPD of 0.8323,0.9037 and 2.75,respectively.(2)For testing the firmness of peach,spectral data were collected at the top,near the equator and bottom positions using a hyperspectral fluorescence imager.PLSR and ELM models were developed and the best acquisition point was found to be near the equator.Modeling of six different pre-processing methods revealed that the GF method performed the best.Five primary feature extraction methods such as particle swarm algorithm(PSO),interval random frog hopping method(IRF)and four secondary feature extraction methods such as IRF-MASS were used to filter the feature variables.Finally,after building extreme learning machine(ELM),multinuclear support vector regression(MK-SVR)and partial least squares regression(PLSR)models for the feature variables extracted by the nine methods,it was found that the GF-VISSA-MASS-MK-SVR method based on the equatorial vicinity achieved the best prediction results with R_p~2,R_c~2and RPD of 0.8882,0.9249 and 3.37,respectively.(3)For p H detection of peach,similarly,hyperspectral fluorescence imager was used to collect spectral data from the top,near the equator and bottom positions of peach.The bottom position was found to be the best acquisition point after comparing the predictive index analysis with PLSR and ELM models.Five different primary pre-processing methods such as moving average smoothing(MA),GF and SG-GF combination methods were used for the bottom position data to attenuate the negative effects such as noise interference and baseline shift in the spectra,and the data results showed that the SG-GF combination method was more effective.Ten different feature extraction methods,including primary and secondary methods such as non-negative matrix decomposition(NMF)and MASS-BOSS,were used to extract the feature variables of the target data.Models such as generalized regression neural network(GRNN)and PSO-LSSVM were constructed to analyze and predict the data by extracting the feature variables.The results show that the SG-FG-BOSS-IRIV-GRNN method based on the bottom position has the best prediction with R_p~2,R_c~2and RPD were0.8840,0.9461 and 3.07 respectively.(4)For the detection of peach sugar at different storage periods(1,3,and 5 days)in 4 environments,spectral data near the equator of peach stored for 1,3,and 5 days were collected using a hyperspectral fluorescence imager.Pre-processing was performed using five different methods,while the three optimal methods from each set of modelling results were combined and used in order to improve the signal-to-noise ratio.The results showed that the best preprocessing method was MA-SG-ES on day 1,MA-GF-ES on day 3,and DT-GF-ES on day 5.Five primary feature extraction methods such as MASS,IRF,and three secondary feature extraction methods such as IRF-BOSS were selected to filter the feature variables.After comparing the prediction results of different models,it was found that the BOSS-MASS-MK-SVR method could be generally applied to the high-precision detection of peach sugar at different storage periods.The reliability of the method was also demonstrated by one-way ANOVA.