| 深度学习在植物表型研究中的应用现状与展望 |
| |
| 引用本文： | 岑海燕,朱月明,孙大伟,翟莉,万亮,麻志宏,刘子毅,何勇.深度学习在植物表型研究中的应用现状与展望[J].农业工程学报,2020,36(9):1-16. |
| |
| 作者姓名： | 岑海燕 朱月明 孙大伟 翟莉 万亮 麻志宏 刘子毅 何勇 |
| |
| 作者单位： | 浙江大学生物系统工程与食品科学学院,杭州 310058;农业农村部光谱检测重点实验室,杭州 310058;浙江大学现代光学仪器国家重点实验室,杭州 310027 |
| |
| 基金项目： | 国家自然科学基金（31971776）；国家重点研发计划课题（2017YFD0201501） |
| |
| 摘 要： | 精确测量植物表型是深入分析表型-基因-环境互作关系,了解植物生理过程的前提和基础,也是培育良种和提升现代农业生产精准管控的关键。伴随高通量植物表型测量与分析技术的不断发展,以深度学习为代表的人工智能方法在植物表型研究与应用中取得了一系列重要进展。为系统阐述相关研究最新成果和热点问题,该研究首先概述了植物表型与深度学习方法的背景;随后从植物识别与分类、胁迫分析、产量预测、面向精准育种和精准管理的表型分析等方面综述了深度学习在植物表型交叉研究的进展;最后提出了未来深度学习和植物表型交叉融合研究与应用中亟需解决的问题,并展望了植物表型研究智能化的发展前景。
|
| 关 键 词： | 植物 表型 管理 深度学习 识别与分类 作物育种 |
| 收稿时间： | 2019/12/18 0:00:00 |
| 修稿时间： | 2020/2/24 0:00:00 |
Current status and future perspective of the application of deep learning in plant phenotype research |
| |
| Cen Haiyan,Zhu Yueming,Sun Dawei,Zhai Li,Wan Liang,Ma Zhihong,Liu Ziyi,He Yong.Current status and future perspective of the application of deep learning in plant phenotype research[J].Transactions of the Chinese Society of Agricultural Engineering,2020,36(9):1-16. |
| |
| Authors: | Cen Haiyan Zhu Yueming Sun Dawei Zhai Li Wan Liang Ma Zhihong Liu Ziyi He Yong |
| |
| Institution: | 1.College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou 310058, China; 2. Key Laboratory of Spectroscopy Sensing, Ministry of Agriculture and Rural Affairs, Hangzhou 310058, China; 3. State Key Laboratory of Modern Optical Instrumentation, Zhejiang University, Hangzhou 310027, China |
| |
| Abstract: | Abstract: Accurate plant phenotyping is important for gaining a fundamental understanding of phenotype-genotype-environment interaction and is also critical for plant breeding and agricultural precision management. With the development of accurate and high-throughput plant phenotyping techniques, big phenotypic data of various plants especially image data can be collected. There is an urgent need to develop effective approaches to dealing with large-scale image data analysis to explore the biological and physiological mechanisms which can be eventually used from the laboratory to the field. This research was entering a new era called ''smart phenomics''. Deep Learning (DL) provided an opportunity to extract useful traits from the complicated phenotypic dataset, which could bridge the knowledge gap between genotype and phenotype for fundamental research and engineering applications in a breeding program and precision farming. Recently, a series of phenotyping related research supported by DL had been published all around the plant fundamental mechanism as well as the agricultural engineering applications. This study investigated the latest publications focused on phenotyping relating to the following algorithms: Convolutional Neural Network (CNN), Restricted Boltzmann Machines (RBM), Auto Encoder (AE), Sparse Coding (SC) and Recurrent Neural Network (RNN), both of the achievements and problems were introduced and summarized in the following aspects. The published researches involved the phenotypic identification and classification over various crops from tissues, organs, and plant scales singly or combined. Not like DBN, SC, or other earlier algorithms, CNN could extract the features without image preprocessing or feature design, its capability also grew rapidly since it was proposed and now had been the first-choice for image identification and classification scenarios. While deep learning applications in biotic/abiotic plant stress analysis mainly focused on the identification and classification of different phenotypic traits of various common crops under typical stresses. Recent studies CNN showed the most potential capability for stress identification and classification, and the predictions by CNN was an irrelevance to the type of stress. Studies also found that the qualitative analysis of abiotic stress could be diagnosed by transfer learning to reduce training time without affecting the prediction capability of the model, especially network architectures with mature applications scenarios manifested stable performance in terms of adaptability and migration based on CNN or integrated with CNN, Besides, yield prediction accuracy had been greatly improved through color, geometric shapes, textures and multiple phenotypic coupling features, which could be divided into the following three scenarios, including fruit yield prediction based on fruit identification and counting, field crop yield prediction by Unmanned Aerial Vehicle (UAV) remote sensing and multi-dimensional yield prediction based on multi-scale, multi-source, and multi-factor data. Moreover, deep learning had shown the potential for precision breeding and precision management. The precise identification of crop-specific phenotypes is the basis for accurate breeding and phenotypic management, and it can be summarized as two categories, including quantitative index counts after target detection and target segmentation under complex field conditions. In summary, a large number of proposed network architectures applied in plant phenotyping have been reviewed, and future efforts should be made on improving the efficiency and accuracy in production scenarios. Finally, the trend and future perspective in the multi-disciplinary research field of deep learning in plant phenotype research were also discussed. |
| |
| Keywords: | plant phenotype management deep learning identification and classification crop breeding |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
| 点击此处可从《农业工程学报》浏览原始摘要信息 |
| 点击此处可从《农业工程学报》下载免费的PDF全文 |
|
