基于RGB图像和CNN模型的水稻氮素诊断系统  

Rice Nitrogen Nutrition Diagnosis System based on RGB Images and CNN Model

作　　者：吕斌[1] 姚强[1] 粟超[1] 李波 查茜 黄祥 詹火木[1] LYU Bin;YAO Qiang;SU Chao;LI Bo;CHA Xi;HUANG Xiang;ZHAN Huomu(Institute of Agricultural Science and Technology Information,Chongqing Academy of Agricultural Sciences,Chongqing 401329,China)

机构地区：[1]重庆市农业科学院农业科技信息研究所,重庆401329

出　　处：《现代农业装备》2024年第4期67-73,共7页Modern Agricultural Equipment

基　　金：重庆市农业发展基金(NKY-2021AB009)。

摘　　要：氮素营养诊断是水稻栽培取得优质高产的关键技术之一。为了能达到利用无人机或智能手机采集水稻冠层图像即可快速获得水稻氮素营养状况和施肥建议处方的研究目的,于2021-2022年在紫色水稻田开展7水平氮素施肥试验,获取不同氮素含量水稻冠层无人机RGB图像,通过深度学习方法获得卷积神经网络(CNN)优化模型,构建基于安卓智能手机的水稻氮素营养诊断系统。结果表明,不同施氮水平对水稻叶片氮素含量的影响,表现在返青期和分蘖期的差异更加明显,共获得10173张RGB冠层图像。通过调整CNN训练参数batch_size、epoch、learning_rate及图像缩放比例,获得准确率超过80%以上的模型9个,开发完成1套基于Android智能手机的APP客户端程序,实现了深度学习模型由Python环境到Android系统的迁移。研究证明,应用无人机或智能手机采集水稻各生长期冠层RGB图像数据,采用深度学习CNN模型构建的基于Android智能手机的水稻氮营养诊断系统,技术方法可行,能够作为水稻生育期氮素营养快速诊断的工具。Nitrogen nutrition diagnosis is one of the key technologies to achieve high yield and quality in rice cultivation.In order to quickly obtain the nitrogen nutrition status and fertilization prescription of rice by using unmanned aerial vehicles(UAVs)or smartphones to collect rice canopy images,a 7-level nitrogen fertilization experiment was conducted in purple rice fields from 2021 to 2022,and rice canopy UAV RGB images with different nitrogen contents were obtained.A convolutional neural network(CNN)optimized model was obtained through deep learning methods to construct a rice nitrogen nutrition diagnosis system based on Android smartphones.Results showed that the impact of different nitrogen levels on rice leaf nitrogen content was more obvious in the tillering and panicle initiation stages.A total of 10173 RGB canopy images were obtained.By adjusting the CNN training parameters such as batch_size,epoch,learning_rate,and image scaling ratio,a total of nine models with accuracy exceeding 80%were obtained.A set of APP client programs based on Android smartphones was developed to achieve the migration of deep learning models from the Python environment to the Android system.By utilizing UAVs or smartphones to collect RGB canopy images of rice at different growth stages,the rice nitrogen nutrition diagnosis system based on Android smartphones constructed by deep learning CNN models is feasible and can be used as a tool for rapidly diagnosing nitrogen nutrition during the rice growth period.

关 键 词：水稻 氮素营养 诊断系统 深度学习 图像识别 无人机 

分 类 号：S126[农业科学—农业基础科学]