摘要
为了解决传统绿豆干旱胁迫识别方法存在识别率低、时效性差的问题,本研究建立了基于卷积神经网络(CNN)和转换器(Transformer)的绿豆干旱胁迫识别模型Mungbean-droughtNet。该模型采用双分支结构,利用全局特征提取模块(GFEM)分支和局部特征提取模块(LFEM)分支分别从输入图像提取局部特征和全局特征。最后利用多层感知器(MLP)模块将局部特征和全局特征进行融合,实现分类。在实际数据分析中,共采集14536张干旱胁迫下的绿豆叶绿素荧光图像,分为HR、R、MR、S、HS和对照6类。利用Mungbean-droughtNet模型对叶绿素荧光图像数据集进行分析,结果表明,Mungbean-droughtNet模型对测试集中叶绿素荧光图像的平均识别准确率为95.57%,平均精度为98.18%,平均召回率为98.40%,平均F1分数为98.28%。和目前先进模型EfficientNetV2和Swin Transformer相比,Mungbean-droughtNet模型准确率分别提高了3.56个百分点和2.62个百分点,表现出更强的鲁棒性和更好的识别效果。本研究结果为绿豆干旱胁迫研究和耐旱基因挖掘提供了基础。
To address the issues of low recognition rate and poor timeliness in traditional methods for identifying drought stress in mung beans,this study established a mung bean drought stress recognition model named Mungbean-droughtNet based on convolutional neural network(CNN) and transformer.The model employed a dual-branch structure,utilized the global feature extraction module(GFEM) branch and the local feature extraction module(LFEM) branch to extract local and global features from the input images,respectively.Finally,multilayer perceptron(MLP) module was used to fuse the local and global features for classification.In the actual data analysis,a total of 14 536 chlorophyll fluorescence images of mung beans under drought stress were collected and classified into six categories,including HR,R,MR,S,HS and the control group.The Mungbean-droughtNet model was applied to analyze the chlorophyll fluorescence image dataset,the results showed that the Mungbean-droughtNet model achieved an average recognition accuracy of 95.57%,an average precision of 98.18%,an average recall ratio of 98.40%,and an average F1-score of 98.28% for the chlorophyll fluorescence images in the test set.Compared with the current advanced models EfficientNetV2 and Swin Transformer,the accuracy of the Mungbean-droughtNet model increased by 3.56 percentage points and 2.62 percentage points,respectively,demonstrating stronger robustness and better recognition performance.This study provides a foundation for research on mung bean drought stress and the excavation of drought-resistant genes.
作者
蒋东山
刘金洋
张浩淼
李士丛
罗仔秋
余骥远
李洁
陈新
袁星星
高尚兵
JIANG Dongshan;LIU Jinyang;ZHANG Haomiao;LI Shicong;LUO Zaiqiu;YU Jiyuan;LI Jie;CHEN Xin;YUAN Xingxing;GAO Shangbing(Faculty of Computer and Software Engineering,Huaiyin Institute of Technology,Huai’an 223003,China;Institute of Economic Crops,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China)
出处
《江苏农业学报》
北大核心
2025年第1期87-100,共14页
Jiangsu Journal of Agricultural Sciences
基金
国家食用豆产业技术体系岗位科学家项目(CARS-09-G13)
江苏省种业揭榜挂帅项目[JBGS(2021)004]
江苏省研究生实践创新计划项目(SJCX24_2146)。
关键词
绿豆
干旱胁迫
卷积神经网络
转换器
图像识别
叶绿素荧光图像
mung bean
drought stress
convolutional neural network
transformer
image recognition
chlorophyll fluorescence image
