基于全透射近红外光谱的空心西瓜在线检测方法研究  

Research on the Method of Online Detection of Hollow Watermelons Based on Full-Transmission Near-Infrared Spectroscopy

作　　者：李佳琪 田喜 王庆艳[2,3] 何鑫 黄文倩 LI Jia-qi;TIAN Xi;WANG Qing-yan;HE Xin;HUANG Wen-qian(Department of Vehicle Engineering,School of Mechanical Engineering,Guangxi University,Nanning 530004,China;Intelligent Equipment Research Center,Intelligent-Meaurment,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;National Research Center of Intelligent Equipment for Agriculture,Intelligent-Meaurment,Beijing 100097,China)

机构地区：[1]广西大学机械工程学院车辆工程系,广西南宁530004 [2]北京市农林科学院智能装备技术研究中心智能检测实验室,北京100097 [3]国家农业智能装备工程技术研究中心智能检测实验室,北京100097

出　　处：《光谱学与光谱分析》2025年第5期1440-1447,共8页Spectroscopy and Spectral Analysis

基　　金：国家重点研发计划课题(2022YFD2002202);国家现代农业产业技术体系西甜瓜岗位科学家项目(CARS-25-2024-G23);北京市农林科学院创新能力建设专项(KJCX20230206)资助。

摘　　要：西瓜具有很高营养价值,医学上具有解暑的功效。成熟度、甜度和是否空心是西瓜评价的关键指标,成为市场竞争力的重要因素,西瓜空心的筛选保证西瓜更高品质,提高市场竞争力。通过实验室自主研发的全透射近红外光谱设备采集307个西瓜光谱。根据西瓜空心位置主要发生在瓜体中心的特点,创新性提出对光谱进行区域分割和权重处理。通过支持向量机(SVM)和偏最小二乘判别分析(PLSDA)算法分别挑选出最优的两种权重光谱,基于原始光谱、权重光谱以及进行多元散射矫正(MSC)和卷积平滑(SGS)预处理后的光谱,3种光谱采用SVM和PLSDA分别进行空心西瓜分类建模。结果显示,相比原始光谱建立的模型,通过预处理并不一定会加强模型效果,甚至会降低模型效果,通过两种权重光谱建立模型效果最好,准确率分别为96.74%(SVM)和92.39%(PLSDA),权重处理后的光谱相比原始光谱和其他两种预处理后的光谱具有更好的建模效果。采用SVM和PLSDA两种算法挑选出的权重光谱和原始光谱分别进行一维卷积神经网络(1D-CNN)建立分类模型,模型准确率分别为98.92%(SVM),96.77%(PLSDA)和95.70%(原始光谱)。结果表明,1D-CNN建模效果相比SVM和PLSDA建模效果更好,并且光谱分割和权重处理后的光谱在1D-CNN中仍然适用,效果相比原始光谱更好,此研究为空心西瓜无损在线分级检测提供了重要的技术支撑。Watermelon has high nutritional value and is known for its effectiveness in relieving heat in medical applications.Key indicators for evaluating watermelon include ripeness,sweetness,and whether it is hollow.These factors significantly influence market competitiveness.Screening for hollow watermelons ensures higher quality,thereby enhancing market competitiveness.In this study,307 watermelon spectra were collected using a fully transmissive near-infrared(NIR)spectroscopy device developed independently in our laboratory.Based on the characteristic that hollow areas in watermelons primarily occur at the center of the fruit,we innovatively propose segmenting and weighting the spectra.The optimal two weighted spectra were selected using Support Vector Machine(SVM)and Partial Least Squares Discriminant Analysis(PLSDA)algorithms.Classification models for hollow watermelons were then built using the original spectra and preprocessed with Multiplicative Scatter Correction(MSC)and Savitzky-Golay Smoothing(SGS)in combination with SVM and PLSDA.The results showed that preprocessing the spectra did not necessarily improve the model performance and could even decrease it compared to models built with the original spectra.The models established using the two weighted spectra achieved the best performance,with accuracies of 96.74%(SVM)and 92.39%(PLSDA).The weighted spectra provided better modeling performance than the original and other preprocessed spectra.The weighted spectra were selected using SVM and PLSDA algorithms,and the original spectra were used to establish classification models with one-dimensional convolutional neural networks(1D-CNN).The model accuracies were 98.92%(SVM),96.77%(PLSDA),and 95.70%(original spectra).The results indicated that 1D-CNN provided better modeling performance than SVM and PLSDA.Additionally,the segmented and weighted spectra remained effective in 1D-CNN and performed better than the original spectra.This study provides important technical support for non-destructive online grading detectio

关 键 词：全透射近红外光谱 空心西瓜 权重光谱 一维卷积神经网络(1D-CNN) 

分 类 号：S375[农业科学—农产品加工]