高光谱和集成学习的鸭梨黑斑病潜育期快速识别方法  

Hyperspectral and Ensemble Learning Method for Rapid Identification of Black Spot in Yali Pear at Gley Stage

作　　者：张凡 王文秀 张宇帆 胡泽轩 赵丹阳 马倩云 石海燕[2] 孙剑锋[1] ZHANG Fan;WANG Wen-xiu;ZHANG Yu-fan;HU Ze-xuan;ZHAO Dan-yang;MA Qian-yun;SHI Hai-yan;SUN Jian-feng(College of Food Science and Technology,Hebei Agricultural University,Baoding 071000,China;College of Horticulture,Hebei Agricultural University,Baoding 071000,China)

机构地区：[1]河北农业大学食品科技学院,河北保定071000 [2]河北农业大学园艺学院,河北保定071000

出　　处：《光谱学与光谱分析》2023年第5期1541-1549,共9页Spectroscopy and Spectral Analysis

基　　金：河北省省属学校基本科研业务费研究项目(KY202002);河北省重点研发项目(20327111D);塔里木大学现代农业工程重点实验室开放课题(TDNG2020102);国家自然科学基金项目(31301761)资助。

摘　　要：鸭梨黑斑病在感染早期阶段引起感染区域外观的变化很微小,肉眼难以观察,因此对其早期识别仍然是困难的。结合高光谱成像技术和Stacking集成学习算法,实现了鸭梨黑斑病的潜育期识别检测。首先,获取健康和不同腐败程度黑斑病鸭梨样品的原始高光谱图像,基于图像选取感兴趣区域(ROI),然后对提取的平均光谱数据进行一阶导数(FD)、二阶导数(SD)、标准正态变量变换(SNVT)及组合SNV-FD和SNV-SD预处理后,采用竞争性自适应权重取样法(CARS)提取特征波长的光谱信息。最后基于筛选出的特征信息分别建立最小二乘支持向量机(LS-SVM)、 K最邻近法(KNN)、随机森林(RF)和线性判别分析(LDA)分类模型。其中,预测效果最好的组合为SNV-FD-LSSVM,SNV-KNN和SNV-FD-RF,准确率分别达到94%, 88%和88%。四种算法建立的模型中,测试集准确率不低于85.00%的个数分别为5、 3、 2和0,因此优选出LS-SVM、 KNN和RF三个分类器用于后续的集成学习。为提高模型准确率,以优选出的LS-SVM、 KNN和RF三种模型作为基分类器构建Stacking学习框架,并与单一分类器建模结果进行对比分析。结果表明,集成学习模型的总体识别正确率达到了98.68%,较单一分类器模型提高了4.64%,且对潜育期样品的识别率提高了11%。证实了高光谱成像结合集成学习方法识别潜育期黑斑病鸭梨样品可行;集成模型显著提高了单一模型的准确性;为鸭梨黑斑病早期检测和病害分级提供一种新的方法,同时为深入研究集成学习算法在光谱定性中的应用奠定了一定基础。It is still difficult to identify black pear spots in the early stage of infection because the changes in the appearance of the infected area are very small and difficult to be observed by the naked eye.This study combined hyperspectral imaging technology and Stacking integrated learning algorithm to realize gley identification and detection of pear black spot.Firstly,a hyperspectral imaging system was used to collect the hyperspectral images of healthy pear samples and different disease grades.The region of interest(ROI)was selected based on the images,and the average spectrum was extracted.Then,First derivative(FD),Second derivative(SD),Standard Normal Variable Transformation(SNVT),SNV-FD and SNV-SD pretreatments were performed on the extracted original spectral data.Then,the Competitive Adaptive Weight Sampling(CARS)method was used to extract the spectral information of the characteristic wavelength.Finally,the Least Square support vector machines(LS-SVM),K-nearest neighbor method(KNN),Random Forest(RF)and Linear discriminant Analysis(LDA)classification models are established respectively based on the screened feature information.Among them,the combination of SNV-FD-LSSVM,SNV-KNN and SNV-FD-RF was better,with test set accuracy of 94%,88%and 88%respectively.In the models established by LS-SVM,KNN,RF and LDA algorithms,the number of test set accuracy not less than 85.00%are 5,3,2 and 0 respectively.Therefore,three classifiers,LS-SVM,KNN and RF,are selected for subsequent ensemble learning.In order to improve the model accuracy,the optimized LS-SVM,KNN and RF models were used as the base classifier to construct the Stacking learning framework,and the modeling results of a single classifier were compared and analyzed.The results showed that the overall recognition accuracy of the integrated learning model is 98.68%,which is 4.64%higher than that of the single classifier model,and the recognition rate of gley samples is 11%higher.The results confirmed the feasibility of hyperspectral imaging combined with an integr

关 键 词：高光谱成像技术 鸭梨黑斑病 Stacking集成模型 潜育期 基模型 

分 类 号：S379.9[农业科学—农产品加工]