深度学习方法在农业信息中的研究进展与应用现状
投稿时间:2019-06-20    点此下载全文
引用本文:傅隆生,宋珍珍,Zhang Xin,李瑞,王东,崔永杰.深度学习方法在农业信息中的研究进展与应用现状[J].中国农业大学学报,2020,25(2):105-120
摘要点击次数: 765
全文下载次数: 325
作者单位E-mail
傅隆生 西北农林科技大学 机械与电子工程学院, 陕西 杨凌 712100
农业农村部农业物联网重点实验室, 陕西 杨凌 712100
陕西省农业信息感知与智能服务重点实验室, 陕西 杨凌 712100
Center for Precision and Automated Agricultural Systems, Washington State University, Prosser, WA 99350, USA 
fulsh@nwafu.edu.cn 
宋珍珍 西北农林科技大学 机械与电子工程学院, 陕西 杨凌 712100  
Zhang Xin Center for Precision and Automated Agricultural Systems, Washington State University, Prosser, WA 99350, USA  
李瑞 西北农林科技大学 机械与电子工程学院, 陕西 杨凌 712100  
王东 西北农林科技大学 机械与电子工程学院, 陕西 杨凌 712100
农业农村部农业物联网重点实验室, 陕西 杨凌 712100
陕西省农业信息感知与智能服务重点实验室, 陕西 杨凌 712100 
 
崔永杰 西北农林科技大学 机械与电子工程学院, 陕西 杨凌 712100
农业农村部农业物联网重点实验室, 陕西 杨凌 712100
陕西省农业信息感知与智能服务重点实验室, 陕西 杨凌 712100 
 
基金项目:陕西省重点研发计划项目(2018TSCXL-NY-05-04,2019ZDLNY02-04);北京工商大学食品安全大数据技术北京市重点实验室开放课题项目(BTBD-2019KF03);国家自然科学基金项目(31971805);西北农林科技大学国际科技合作种子基金项目(A213021803);陕西省青年杰出人才基金项目(Z111021901)
中文摘要:为使农业信息领域的研究人员能够系统和快速地了解深度学习在农业中的研究进展以及应用现状,对深度学习在农业信息领域的应用进行归纳、梳理、分析和展望。对涉及农业领域且应用深度学习技术的90项研究中所涉及的农业问题、具体模型和框架、数据集的来源和特征以及预处理方法、模型评价指标等进行归纳总结分析,并讨论深度学习的优点和局限性,进而展望深度学习的发展趋势。农业领域中的应用包括作物及其器官分类、病虫害识别、果实识别和计数、植物识别、土壤覆盖分类、杂草识别、行为识别和分类、植物养分含量估计、植物叶片或种子表型分析等方面;大多数研究采用卷积神经网络,如AlexNet、VGG16和Faster R-CNN。在框架方面,Caffe使用频次最高,其次是Tensorflow和Keras/Theano;分类准确度是最常用的模型评价指标,其次是F1得分和平均精度。与其他常用方法和技术相比,深度学习不仅精度高,而且性能优于现有的常用图像处理技术。其他涉及计算机视觉技术的农业应用有望通过深度学习技术的使用获得更好的效果。
中文关键词:深度学习  卷积神经网络  应用框架  评价指标  数据增广  迁移学习
 
Applications and research progress of deep learning in agricultureFU Longsheng1, 2, 3, 4*, SONG Zhenzhen1, ZHANG Xin4, LI Rui1, WANG Dong1, 2, 3, CUI Yongjie1, 2, 3
Abstract:The application of deep learning(DL)in agriculture was summarized, organized, analyzed and forecasted in this study, so that researchers in this field could more systematically and quickly understand the application and research progress of DL in agriculture. A total of 90 publications that applied DL in agriculture from 2013 to 2019 were comprehensively reviewed. In addition, the employed models and frameworks, data augmentations, and overall evaluation of the performances were examined. All reviewed publications were categorized into 13 research areas, including the crops and organs classification(11 papers), pest identification(15 papers), fruit detection and counting(13 papers), plant recognition(6 papers), land cover classification(9 papers), identification of weeds(7 papers), behavior identification and classification(7 papers), estimation of plant nutrient content(5 papers), and plant leaf and seed phenotyping(7 papers). It is noticeable that almost all the studies(87 out of 90)were published during or after 2015, indicating a strong tendency of DL applications in agricultural field. More specifically, there were 15, 39, 21, 9 and 6 published in 2019, 2018, 2017, 2016 and 2015, respectively. Most of the studies adopted the architectures of convolutional neural networks(CNNs), such as AlexNet, VGG16 and Faster R-CNN. In terms of the framework, Caffe was the most favored by researchers, followed by Tensorflow and Keras/Theano. A possible reason was that Caffe incorporated various CNN frameworks and datasets, which could be easily adapted for multiple tasks. Besides, it was worth mentioning that most of the works could benefit from transfer learning(i. e. pre-trained networks)concerning leveraging the already existing algorithm parameters of related tasks in order to increase the learning efficiency of the modified model. Some works augmented the dataset to artificially enlarge the number of training images. The major methods of data transformation included image rotating, cropping, scaling, transposing, and mirroring. Regarding the metrics used for performance evaluation, classification accuracy(CA)was the most popular metric used, followed by F1 score and average precision(AP). As reported, the DL models outperform than other approaches(e. g. traditional machine learning)implemented for comparison purposes. CNNs based DL showed 1%-9% higher CA in comparison to support vector machines(SVM), and was superior than unsupervised feature learning with 3%-13% higher CA, as well as 2%-46% greater CA in relation to local shape and color features. Moreover, DL showed advantages in searching feature engineering(e. g. good feature extractions), which was not always an easy task being completed manually. The DL models performed robustly even under challenging conditions, such as various illuminations, complex backgrounds, different resolution, size and orientation of the images. Although DL normally costed longer time to train the dataset than other approaches, it had much higher efficiencies on testing data. However, a considerable drawback and barrier in the use of DL was the need of large datasets, which served as the input during the training procedure. Another limitation was that the DL models could learn some particular problems well, but they couldn't generalize beyond the dataset. In the future, to achieve the smarter and more sustainable farming, as well as more secure food production, the DL is expected to be applied in other areas of agriculture where it has not been adequately used yet.
keywords:deep learning  convolutional neural network  application framework  evaluation metrics  data augmentation  transfer learning
查看全文  查看/发表评论  下载PDF阅读器