摘要
为实现棉花生育时期的快速识别,本研究将无人机遥感与图像识别和深度学习技术相结合,以不同品种、不同种植密度的棉花为研究对象,基于搭载RGB相机的无人机获取棉花不同生育时期的冠层图像,对获取的图像进行分割,并将分割后的图像数据集作为输入向量训练卷积神经网络模型(VGGNet19和GoogleNet),建立识别棉花关键生育时期的模型,然后通过对比分析2个学习率(0.0001和0.0005)和4个dropout(0.4、0.5、0.6、0.7)组成的8个训练模型,筛选出最优模型。结果表明,GoogleNet训练模型在学习率为0.0001、dropout为0.6时的准确率最高(98.06%),性能评估指标最好;而VGGNet19训练模型在学习率为0.0005、dropout为0.5时的准确率最高(72.26%),评估效果最好。综合对比模型训练结果、训练轮数、预测集测试结果以及性能评价,GoogleNet卷积神经网络(CNN)模型对棉花关键生育时期识别的效果优于VGGNet19模型,且学习率为0.0001、dropout为0.6时的GoogleNet训练模型是识别棉花关键生育时期的最优模型。本研究为棉花生育时期的快速识别提供了技术支撑,也为其他农作物的生育期研究提供了借鉴方法。
To achieve rapid identification of cotton growth stages,this study combined unmanned aerial vehicle(UAV)remote sensing with image recognition and deep learning technology,different varieties and planting densities of cotton are taken as the research subjects.UAVs equipped with RGB cameras were used to capture canopy images of cotton at different growth stages.These images were segmented,and the segmented image dataset was used as input vectors to train convolutional neural network models(VGGNet19 and GoogleNet),establishing a model for identifying key growth stages of cotton.Then,by comparing and analyzing 8 training models composed of 2 learning rates(0.0001 and 0.0005)and 4 dropout rates(0.4,0.5,0.6,0.7),the optimal model was selected.The results showed that the GoogleNet training model had the highest accuracy(98.06%)and the best performance evaluation metrics when the learning rate was 0.0001 and the dropout rate was 0.6,while the VGGNet19 training model had the highest accuracy(72.26%)and the best performance evaluation metrics when the learning rate was 0.0005 and the dropout rate was 0.5.By comprehensively comparing model training results,training epochs,prediction set test results,and performance evaluation,the GoogleNet convolutional neural network(CNN)model performed better in identifying key growth stages of cotton than the VGGNet19 model.The GoogleNet model trained with a learning rate of 0.0001 and a dropout rate of 0.6 was the optimal model for identifying key growth stages of cotton.This study provided technical support for the rapid identification of cotton growth stages and offered a reference method for the research of other crops.
作者
李明泽
张静
雷亚平
韩迎春
王国平
陈国栋
李亚兵
冯璐
LI Mingze;ZHANG Jing;LEI Yaping;HAN Yingchun;WANG Guoping;CHEN Guodong;LI Yabing;FENG Lu(College of Agriculture,Tarim University,Alar,Xinjiang 843300;Cotton Research Institute,Chinese Academy of Agricultural Sciences,Anyang,Henan 455000)
出处
《核农学报》
CAS
CSCD
北大核心
2024年第10期2020-2031,共12页
Journal of Nuclear Agricultural Sciences
基金
河南省科技攻关项目(232102110030)
中央级公益性科研院所基本科研业务费专项(1610162023004)。
关键词
卷积神经网络
深度学习
无人机
农艺性状
产量
convolutional neural network
deep learning
unmanned aerial vehicle
agronomic traits
production
