基于高光谱成像技术的涌泉蜜桔糖度最优检测位置  

Optimal position for suger content detection of Yongquan honey oranges based on hyperspectral imaging technology

作　　者：李斌[1] 万霞 刘爱伦 邹吉平 卢英俊 姚迟 刘燕德[1] LI Bin;WAN Xia;LIU Ai-lun;ZOU Ji-ping;LU Ying-jun;YAO Chi;LIU Yan-de(Intelligent Electromechanical Equipment Innovation Research Institute,East China Jiaotong University,National and Local Joint Engineering Research Center of Fruit Intelligent Photoelectric Detection Technology and Equipment,Nanchang 330013,China)

机构地区：[1]华东交通大学、智能机电装备创新研究院、水果智能光电检测技术与设备国家与地方联合工程研究中心,南昌330013

出　　处：《中国光学（中英文）》2024年第1期128-139,共12页Chinese Optics

基　　金：青年科学基金项目(No.12103019)。

摘　　要：本文旨在探索涌泉蜜桔糖度的最优检测位置和最佳预测模型,以便为蜜桔糖度检测分级提供理论依据。本文利用波长为390.2~981.3 nm的高光谱成像系统对涌泉蜜桔糖度最佳检测位置进行研究,将涌泉蜜桔的花萼、果茎、赤道和全局的光谱信息与其对应部位的糖度结合,建立其预测模型。使用标准正态变量变换(SNV)、多元散射校正(MSC)、基线校准(Baseline)和SG平滑(Savitzkv-Golay)4种预处理方法对不同部位的原始光谱进行预处理,用预处理后的光谱数据建立偏最小二乘回归(PLSR)和最小二乘支持向量机(LSSVM)模型。找出蜜桔不同部位的最佳预处理方式,对经过最佳预处理后的光谱数据采用竞争性自适应重加权算法(CARS)和无信息变量消除法(UVE)进行特征波长筛选。最后,用筛选后的光谱数据建立PLSR和LSSVM模型并进行分析比较。研究结果表明,全局的MSC-CARS-LSSVM模型预测效果最佳,其预测集相关系数Rp=0.955,均方根误差RMSEP=0.395,其次是蜜桔赤道部位的SNV-PLSR模型,其预测集相关系数Rp=0.936,均方根误差RMSEP=0.37。两者预测集相关系数相近,因此可将赤道位置作为蜜桔糖度的最优检测位置。本研究表明根据蜜桔不同部位建立的糖度预测模型的预测效果有所差异,研究最优检测位置和最佳预测模型可以为蜜桔进行糖度检测分级提供理论依据。The objective of this study is to explore the optimal detection location and the best prediction model of the suger level of Yongquan honey oranges,which can provide a theoretical basis for the brix measurement and classification of honey oranges.With the wavelength range of 390.2−981.3 nm hyperspectral imaging system was used to study the best position for detecting the sugar content of Yongquan honey oranges,and the spectral information of the calyx,fruit stem,equator and global of Yongquan honey oranges were combined with their sugar content of corresponding parts to establish its prediction model.The original spectra from the different locations were pre-processed by Standard Normal Variance(SNV)transformation,Multiple Scattering Correction(MSC),baseline calibration(Baseline)and SG smoothing,respectively,and the Partial Least Squares Regression(PLSR)and Least Squares Support Vector Machine(LSSVM)models were established based on the pre-processed spectral data.The best pre-processing methods for different parts of the honey oranges were found,and the optimal spectral data obtained by the best pre-processing methods were conducted to identify characteristic wavelengths using the Competitive Adaptive Re-weighting Sampling algorithm(CARS)and Uninformative Variable Elimination(UVE).Finally,the PLSR and LSS-VM models were established and compared based on the selected spectral data.The results show that the global MSC-CARS-LSSVM model demonstrates the most accurate prediction performance,with a correla-tion coefficient of Rp=0.955 and an RMSEP value of 0.395.Alternatively,the SNV-PLSR model of the equatorial location of honey oranges was found to be the next more effective,with a correlation coefficient of Rp=0.936,and an RMSEP value of 0.37.The correlation coefficients of the two prediction models are simil-ar,the equatorial location can be used as the optimal position for measuring the sugar content of honey or-anges.This study demonstrates that the prediction models based on different parts of the orange have

关 键 词：涌泉蜜桔 高光谱 糖度 偏最小二乘回归 最小二乘支持向量机 

分 类 号：O433.4[机械工程—光学工程]