高光谱结合深度学习的桑椹采后TSS含量无损检测  

Nondestructive Determination of TSS Content in Postharvest Mulberry Fruits Using Hyperspectral Imaging and Deep Learning

作　　者：王子轩 杨良 黄凌霞 何勇 赵丽华 占鹏飞 WANG Zi-xuan;YANG Liang;HUANG Ling-xia;HE Yong;ZHAO Li-hua;ZHAN Peng-fei(Business School,Hohai University,Nanjing 210024,China;Key Laboratory of Silkworm and Bee Resource Utilization and Innovation of Zhejiang Province,Hangzhou 310058,China;Institute of Sericulture,Huzhou Academy of Agricultural Sciences,Huzhou 313002,China;Key Laboratory of Spectroscopy Sensing,Ministry of Agriculture and Rural Affairs,Hangzhou 310058,China)

机构地区：[1]河海大学商学院,江苏南京210024 [2]浙江省蚕蜂资源利用与创新研究重点实验室,浙江杭州310058 [3]湖州市农业科学研究院蚕桑所,浙江湖州313002 [4]农业农村部光谱检测重点实验室,浙江杭州310058

出　　处：《光谱学与光谱分析》2024年第6期1724-1730,共7页Spectroscopy and Spectral Analysis

基　　金：湖州市重点研发计划项目(2022ZD2053);浙江省农业重大技术协同推广项目(2022XTTGCS03)资助。

摘　　要：桑椹起源于中国,是我国最具历史的“药食同源”水果之一。但桑椹采后快速变质和皮薄易腐败的特点制约了其产业化发展。总可溶性固形物(TSS)是决定桑椹风味和品质的重要成分,是其商业化的最基本品质特性之一。借助于近红外高光谱成像技术和深度学习方法优化桑椹采后TSS含量的预测模型,同时评估采后常见储运温度条件对定量模型的影响,为桑椹采后品质快速评价提供依据。选用具有一致商业成熟度的桑椹分别在常温(25℃)和低温(4℃)储藏,然后在不同储藏阶段对样本进行光谱数据采集和TSS含量测定直至桑椹腐败不适宜食用。基于校正后的高光谱图像提供的空间信息提取感兴趣区域以获得无背景的代表性光谱,然后将标准正态变换(SNV)、多元散射校正(MSC)、Savizkg-Golag(SG)平滑用于光谱的预处理,以提升光谱信噪比。利用深度学习方法实现了桑椹采后TSS含量的预测。对于常温和低温桑椹样本,最优CNN模型剩余预测偏差(RPD)值分别达到5.828和5.449,预测均方根误差(RMSEP)值分别为1.082和1.099°Brix,可见低温条件储藏降低了CNN模型的预测性能。为进一步验证CNN模型的效果,建立了基于传统经典机器学习方法偏最小二乘(PLS)和最小二乘支持向量机(LS-SVM)的TSS含量预测模型。结果表明,非线性模型LS-SVM更适合桑椹的TSS含量预测。对于两个不同储藏温度,最优LS-SVM模型RPD值分别为4.221和4.423,表明CNN优于经典机器学习方法。综上所述,高光谱成像结合深度学习CNN的桑椹采后TSS预测具有较大潜力,这为桑椹品质快速检测提供了技术支撑。Originating in China,mulberry is one of the fruits of the homology of medicine and food and has a long history.However,the industrialization of mulberry fruit has been limited by its characteristics of short maturity period and the tendency for thin skin to decay.Total soluble solid(TSS)is an important component of determining the mulberry flavor and qualityandis one of the most basic quality characteristics for its postharvest-commercialization.This study aims to optimize a prediction model for monitoring the TSS content in postharvest mulberry fruits using near-infrared hyperspectral imaging and deep learning methods and to evaluate the impact of common postharvest storage temperature on the quantitative models,thus providing support for rapid quality assessment of mulberry fruits.Mulberry fruits with consistent commercial maturity were selected for storage at room temperature(25℃)and low temperature(4℃).Samples from different storage stages were selected for spectral data collection and TSS content determination until mulberry fruits became unfit for consumption.Based on the spatial information provided by the corrected hyperspectral images,regions of interest were extracted to obtain representative spectra without background accurately.Then,standard normal variate(SNV),multiplicative scatter correction(MSC),and Savizkg-Golag(SG)smoothing were used for spectra preprocessing to improve the spectral signal-to-noise ratio.Prediction models for TSS content measurement in postharvest mulberry fruits were established using deep learning.For mulberry samples stored at room temperature and low temperature,the optimal CNN models obtained the residual prediction deviation values of 5.828 and 5.429,with the root mean square error of prediction(RMSEP)values of 1.082 and 1.099°Brix,respectively,indicating that the prediction performance of the CNN model was degraded due to the low-temperature storage.The classical machine learning methods of partial least squares(PLS)and least square support vector machine(LS-SVM)were

关 键 词：桑椹 采后 总可溶性固形物 高光谱成像 深度学习 卷积神经网络 

分 类 号：S888[农业科学—特种经济动物饲养]