基于深度学习、小波变换和可见光谱的茶树冻害程度评估  

Evaluation of Freezing Injury Degree of Tea Plant Based on Deep Learning,Wavelet Transform and Visible Spectrum

作　　者：李赫 王玉[2] 范凯[2] 毛艺霖 丁仕波 宋大鹏 王梦琪 丁兆堂 LI He;WANG Yu;FAN Kai;MAO Yi-lin;DING Shi-bo;SONG Da-peng;WANG Meng-qi;DING Zhao-tang(Tea Research Institute of Shandong Academy of Agricultural Sciences,Jinan 250108,China;Tea Research Institute of Qingdao Agricultural University,Qingdao 266109,China;Tea Research Institute of Rizhao Academy of Agricultural Sciences,Rizhao 276827,China)

机构地区：[1]山东省农业科学院茶叶研究所,山东济南250108 [2]青岛农业大学茶叶研究所,山东青岛266109 [3]日照市农业科学研究院茶叶研究所,山东日照276827

出　　处：《光谱学与光谱分析》2024年第1期234-240,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(32002087);山东省农业科技创新重大应用项目(SD2019ZZ010);山东省现代农业产业技术体系基金项目(SDAIT-10-01);山东省泰山学者专项基金项目(ts201712057)资助。

摘　　要：茶树冻害鉴定是评价茶树抗逆性、指导茶园越冬管理的基础。茶树冻害鉴定的传统方法主要通过人工观察茶树叶片的冻害数量和冻害程度,存在准确率低,效率低,主观性强等缺点。提出一种基于深度学习、小波变换和可见光谱的茶树冻害程度评估框架。首先,采集了1 000张茶树受冻后的树冠图像,并且按照4∶1分为训练集和测试集,并对训练集图像的冻害叶片进行标注。其次,采用Faster R-CNN网络对茶树的冻害叶片进行识别提取,并且分别选择了AlexNet、 VGG19和ResNet50三种特征提取器,选择鲁棒性最高的特征提取器作为主干网络。然后,将提取到的茶树冻害叶片图像进行小波变换增强处理,从而得到了一张低频率和三张高频率的图像。将小波变换处理后的图像和未经小波变换处理的图像分别输入到VGG16、 SVM、 AlexNet、 ResNet50等网络对冻害叶片进行分级,比较四种网络的分类性能。根据冻害叶片的数量、冻害叶片的程度,不同冻害程度叶片的权重系数,对茶树受冻程度打分,从而对茶树整体的冻害程度进行评估。结果表明:(1)基于ResNet50的Faster R-CNN模型提取茶树冻害叶片的性能最好,其查准率为93.33%,查全率为92.57%,高于VGG19和AlexNet作为主干网络的识别性能,可以确保大多数的冻害叶片都能被提取出来,并为叶片的冻害程度进一步分级提供基础。(2)VGG16模型分类不同冻害程度的叶片的整体准确率为89%,高于其他模型(SVM、 AlexNet、 ResNet50)的整体分类准确率,表明VGG16模型具有较高的鲁棒性。(3)小波变换处理后的冻害叶片图像与未经小波变换处理后的冻害叶片相比较,能提升模型的整体分类准确率2%~6%。说明小波变换增强技术可以提高网络的准确率。因此,本实验框架可以准确和高效地将茶树冻害叶片进行分级,对于茶园冻害程度的评估具有重要价值,为北方茶园的越冬防护�Identifying the freezing injury of tea plants is the basis of evaluating the stress resistance of tea plants and guiding the overwintering management of tea plantations.The traditional method of tea plant freezing injury identification observes the number and degree of freezing injury of tea leaves manually,which has some disadvantages,such as low accuracy,low efficiency,strong subjectivity and so on.A framework for evaluating freezing injury of tea trees based on deep learning,wavelet transform,and visible spectrum is proposed.Firstly,we collected 1000 crown images of frozen tea trees,divided them into the training sets and test sets according to 4∶1,and labeled the frozen leaves in the training set images.Secondly,we use the Faster R-CNN network to identify and extract tea plants'frozen leaves,select three feature extractors:AlexNet,VGG19 and ResNet50 respectively,and select the feature extractor with the highest robustness as the backbone network.Then,the extracted image of frozen leaves of tea plants is enhanced by wavelet transform,and one low-frequency image and three high-frequency images are obtained.Then the images processed by wavelet transform and those not processed by wavelet transform are input into VGG16,SVM,AlexNet,ResNet50 and other networks to classify the frozen leaves,and the classification performance of the four networks is compared.Finally,according to the number of frozen damaged leaves,the degree of frozen damaged leaves and the weight coefficient of leaves with different degrees of frozen damage,the freezing degree of tea plants is scored to evaluate the overall freezing degree of tea plants.The results show that:(1)the Faster R-CNN model based on ResNet50 has the best performance in extracting frozen leaves of tea plants,with a precision rate of 93.33%and a recall rate of 92.57%,which is higher than the recognition performance of VGG19 and AlexNet as the backbone network,which can ensure that most frozen leaves can be extracted and provide a basis for further classification of the deg

关 键 词：茶树 冻害叶片 Faster R-CNN VGG16 小波变换 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]