基于反射率、吸光度和Kubelka-Munk光谱的贡梨不同损伤程度检测  

Detection of Different Levels of Damage in Gong Pears Based on Reflectance/Absorbance/Kubelka-Munk Spectroscopy

作　　者：李斌[1,2] 卢英俊 苏成涛 刘燕德 LI Bin;LU Ying-jun;SU Cheng-tao;LIU Yan-de(School of Intelligent Electromechanical Equipment Innovation Research Institute,East China Jiaotong University,Nanchang 330013,China;School of Mechatronics and Vehicle Engineering,East China Jiaotong University,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,江西南昌330013 [2]华东交通大学机电与车辆工程学院,江西南昌330013

出　　处：《光谱学与光谱分析》2024年第11期3101-3108,共8页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金青年科学基金项目(12103019)资助。

摘　　要：贡梨在收获、运输和销售过程中容易发生机械损伤,加速果实腐烂,降低贡梨品质。快速判别贡梨的不同损伤程度,进而采取不同的处理措施,以降低经济损失。以往运用高光谱技术研究水果的损伤程度,通常仅用反射率光谱。该研究运用高光谱技术获取贡梨的反射率(R)、吸光度(A)、和Kubelka-Munk(K-M)变换光谱结合3种深度学习算法对健康和不同损伤程度的贡梨进行判别。首先,选取60个新鲜无损伤贡梨作为健康样品,再利用自由落体碰撞装置制备出Ⅰ级损伤、Ⅱ级损伤、Ⅲ级损伤贡梨样品各60个。通过高光谱成像系统采集这240个贡梨样品的光谱数据,对采集的光谱进行黑白校正,以获得贡梨的反射率(R)、吸光度(A)、和Kubelka-Munk(K-M)变换光谱,然后用基准线校准(Baseline)、去趋势(De-Trending)、移动平均(MA-S)、乘法散射校正(MSC)、卷积平滑(SG-S)、标准正态变量变换(SNV)共6种预处理方法对3种原始光谱数据进行预处理,并建立BP神经网络(BP)、极限梯度提升(XGBoost)和随机森林(RF)判别分析模型对贡梨不同损伤程度进行判别。根据模型对贡梨损伤程度的判别结果显示,基于反射率、吸光度、K-M变换光谱的BP模型判别准确率较好,整体准确率达到了85%及以上,且发现经过Baseline预处理后的反射率光谱建立的BP模型比未经预处理的反射率光谱谱建立的BP模型有较大的提升,判别准确率达到了93.33%。为了提升BP模型的精准度和运行效率,对3种原始光谱和Baseline预处理后的光谱利用竞争性自适应重加权(CARS)和无信息变量消除(UVE)方法筛选出特征波段光谱信息,用筛选后的特征光谱数据来建立BP模型,其判别结果显示A-RAW-CARS-BP模型具有最佳的判别准确率,整体准确率达到了96.66%。结果表明,采用3种原始光谱对贡梨的损伤程度进行判别具有可行性,为高光谱技术检测贡梨的不同损伤程度提供了�Gong pear is prone to mechanical damage during harvesting,transportation,and sales,accelerating fruit decay and reducing its quality.Different treatment measures were taken to quickly distinguish the different degrees of damage to the gong pear and reduce economic loss.In the past,hyperspectral technology was used to study the damage degree of fruits.Usually,only the reflectance spectrum was used for thestudy.In this study,the reflectance(R),absorbance(A),and Kubelka-Munk(K-M)spectra of Gong pears were obtained by hyperspectral technology and combined with three deep learning algorithms to distinguish healthy and different damage degrees of Gong pears.Firstly,60 fresh and undamaged Gong pears were selected as healthy samples,and 60 samples ofⅠ,Ⅱ,andⅢdamaged Gong pears were prepared by free fall collision device.The spectral data of these 240 Gong pear samples were collected by hyperspectral imaging system,and the acquired spectra were corrected in black and white to obtain the original spectra of reflectance(R),absorbance(A),and Kubelka-Munk(K-M)of Gong pear.Then,three kinds of original spectral data were preprocessed by Baseline calibration,De-Trending,moving average(MA-S),multiple-scattering correction(MSC),convolution smoothing(SG-S),and standard normal variable transformation(SNV).BP neural network(BP),Limit gradient lift(XGBoost),and random forest(RF)discriminant analysis models were established to distinguish different damage degrees of Gongli.According to the discrimination results of the model on the damage degree of Gong pear,the accuracy of the BP model based on reflectance,absorbance,and K-M spectrum is better,with the overall accuracy reaching 85%or more.Itwas found that the BP model established by the baseline reflectance spectrum after pretreatment showed a greater improvement than that established by the unpretreated reflectance spectrum.The accuracy of discrimination reached 93.33%.To improve the accuracy and operation efficiency of the BP model,competitive adaptive reweighting(CARS)and no-i

关 键 词：贡梨 高光谱 损伤程度 反射率光谱 吸光度光谱 Kubelka-Munk光谱 

分 类 号：O433.4[机械工程—光学工程]