近红外光谱结合深度学习的苹果糖心区域面积占比预测

基金项目：兴滇英才支持计划项目（YNWR-QNBJ-2018-349）；云南省科技厅创新引导与科技型企业培育计划项目（202204BP090005、202304BU090015）

作者简介：马亚星（2001-），男，河南平顶山人，硕士研究生，主要从事农产品无损检测方面的研究。（E-mail）1912832237@qq.com

Abstract： The amount of watercore region in apples hasacertainimpact on their taste and sellingprice，but the internal watercoreofapplesstored foralong timewill graduallydisappear，even affecting the quality of apples aswell astheir selling price.In order to investigate the correlation between near-infrared spectral data and the area of the watercoreregionofapples，andtoestablishapredictionmodel for the proportion of watercore area，watercore apples in the

mature stage were taken as theresearchobjects，andtheirspectral data were colected.By slicing the appes，deeplearning algorithmwasusedtoextracttheregionofthewatercoreineachsection，anddeterminedtheratiooftheareaof theextracted watercoreareaofachcrossectiontotheaeaofteoleross-setionasteproportionofwatercoreareaofthatcoss-section，andthe maximum valueof theproportionofwatercore areawas taken as the measurement valueof theoverallproportion of watercorearea.Theoriginal spectraldata were pretreated with multiplescateringcorection（MSC），standard normalvariate（SNV），standardization，etc.，and partialleast squares regresion （PLSR），support vector machine（SVM）andrandom forest（RF）machine learning prediction models，aswellasconvolutional neutral networks（CNN）andbidirectional long short-term memory（BiLSTM）deeplearning prediction models were established，among which the CNN model establishedusing spectraldataafterstandardized pretreatmenthadthebest efect.Inorder tofurtherimprovethemodeling efect，feature extractionmethodssuchasuninformativevariables elimination（UVE），successve projections algorithm（SPA）andcompetitiveadaptive reweighted sampling（CARS）were used to optimize thedataafter preprocesing and to compare the modeling effct between diferentfeature extractionmethods.Theresultsshowedthatthe prediction modelof apple watercorearea proportion constructed by CNN convolutional neural network was beter than the traditional machine learning modeling.The CNN model constructed by UVE + CARS combined feature extraction algorithm after standardized preprocessing had the best prediction effect，and the determination coefficient （ R² ）and root mean squared error（RMSE）of the validation set were O.921 and 1.882，respectively.Thisstudyprovesthatthepredictionmodelofapplewatercoreareaproportionestablishedbynear-infrared spectroscopycombined with CNNconvolutional neural network can beter predict the proportionof watercore areaof apples， and it provides the theoretical support for nondestructive testing of watercore apples.

Key Words： apple；watercore area proportion；near-infrared spectrum；deep learning； prediction

水果具有的重要营养价值，已经成为人们生活中的必需品。苹果已成为中国主要水果中产量占比较大的一种水果。随着苹果产量不断增加，苹果的种类也随之增多。糖心苹果作为苹果的一种，多生长在昼夜温差大、光照充足的地方[1-]。由于糖心苹果在生长过程中，果实细胞间会形成一种半透明的水渍状物质并沿苹果心室由内向外扩展，即形成“水心”，也叫做\"糖心”[3]，因此其与普通苹果相比口感更佳，深受消费者的喜爱。糖心区域适中，不仅可以提升苹果口感，也可以提高苹果价格，但是，随着苹果长时间的储存，糖心也会逐渐消失，甚至在苹果内部出现腐烂的现象，导致苹果可食性降低[4]。因此，找到一种合适的方法对苹果糖心区域进行检测，对于糖心苹果的储存、售卖及苹果的经济价值提升具有重要意义。

以往对于苹果糖心区域的检测主要依靠人工肉眼判断，童盼盼等1将苹果切成两半，观察剖面的糖心散射半径，对糖心区域进行判断。但是，由于糖心分布不规则，导致这种方法得到的结果不准确，而且耗费时间。近年来，随着人们对于水果品质要求的提高，越来越多的检测方法随之出现[5-7]。其中近红外光谱技术和高光谱技术以其快速无损、操作简单、无化学污染等特点[8-9]，成为水果质量检测的重要手段[0-11]。……