-
草地贪夜蛾(Spodoptera frugiperda)是一种危害性极强的鳞翅目害虫,起源于美洲,具有很强的迁飞能力[1].草地贪夜蛾的幼虫以禾本科、豆科等农作物为食,尤为喜食嫩叶,对农作物危害极大.根据草地贪夜蛾对玉米和水稻的偏好性,将其基因型分成了“玉米型”(Corn-Strain)和“水稻型”(Rice-strain)[2].两种亚型的草地贪夜蛾尽管对食物的偏好性有所不同,但它们具有相同的形态学特征[3].
草地贪夜蛾于2019年初入侵我国云南地区,对当地的农作物造成了严重危害[4].郭井菲等[5]通过测序分析了侵入我国草地贪夜蛾的基因组,发现侵入我国的草地贪夜蛾主要为“玉米型”,也存在着少量的“水稻型”(Rice-strain).草地贪夜蛾能在(11~30) ℃的环境下生长,随着温度的逐渐升高,草地贪夜蛾开始大范围迁移.目前,短短半年时间已扩散蔓延至我国20个省的1 128多个县(市、区),受害面积约56.7多万hm2.另据2019年7月5日日本共同社报道,草地贪夜蛾已入侵日本,在日本鹿儿岛已经发现草地贪夜蛾.
如何快速识别草地贪夜蛾,对尽早防治草地贪夜蛾有着至关重要的作用.深度学习[6-9]的出现推动了人工智能新的浪潮,其优势在图像识别领域发挥得淋漓尽致.深度学习的手段依靠深度卷积神经网络[7, 10-12]逐层提取图片的高级特征,并利用特有的多任务分类器Softmax进行分类,对于图像识别具有很好的效果.因此,本文首先建立包含不同地域草地贪夜蛾的图像数据库,然后根据草地贪夜蛾幼虫头部均具有“Y”字纹,尾部有4个呈正方形排列黑点的特佂[13],设计了三通道T型深度卷积神经网络T-CNN(Convolutional Neural Network,CNN)来自动识别草地贪夜蛾,识别率达到97%.本文提出的识别算法有助于对草地贪夜蛾进行快速准确地识别,为草地贪夜蛾防控和预报提供了重要的技术储备.
A CNN-Based Automatic Identification System for Spodoptera frugiperda
-
摘要: 草地贪夜蛾是一种严重破坏农作物的重大洲际害虫,对我国农业生产造成了极大的威胁.尽管一系列防治措施已经展开,但如何有效辨别草地贪夜蛾仍然是防控工作中的一大难题.为了建立一个有效的识别算法,课题组开展了一系列研究工作,主要贡献在于:①采集了不同地域、不同生长区间的草地贪夜蛾及相似物种图片,建立了一个草地贪夜蛾识别数据库;②利用基于特征融合的深度学习算法,建立了一个三通道T型深度卷积神经网络(T-CNN),在现有数据集上平均识别率达到97%,为草地贪夜蛾的智能识别与防控工作提供了技术支撑.Abstract: Spodoptera frugiperda is a serious crop-destroying pest, which poses a great threat to agricultural production in China. Although a series of preventive measures have been adopted, how to identify the pest effectively is still a major problem in the field. In a study reported in this paper, a series of work was done to establish an effective recognition algorithm. Our main contributions were as follows. First, pictures of S. frugiperda and similar species were collected from different regions, and a recognition database of S. frugiperda was established. Secondly, using a deep-learning algorithm based on feature fusion, we constructed a three-channel T-type deep convolution neural network (T-CNN), whose average recognition rate was over 97% on the existing data sets, thus providing technical support for the smart identification and control of S. frugiperda.
-
Key words:
- Spodoptera frugiperda /
- recognition /
- deep learning /
- convolutional neural networks (CNN) /
- feature fusion .
-
表 1 用于建立草地贪夜蛾形态数据库的图片内容及拍摄来源
巫山 江津 云阳 FAO 总数 虫卵 4 4 幼虫 24 3 2 2 31 蛹 1 1 成虫 2 2 总数 25 3 2 8 38 表 2 拍摄数据集划分
种类 卵 雌成虫 雄成虫 巫山幼虫 江津幼虫 云阳幼虫 FAO幼虫 非草地贪夜蛾 数量 30 30 30 30 30 30 30 30 表 3 T-CNN网络参数设置
T-CNN的网络参数 通道C1 通道C2 通道C3 网络层 网络层参数 网络层 网络层参数 网络层 网络层参数 Input1 (64,64) Input2 (64,64) Input3 (1 764,1) Conv1 (96,5,5,2) Flatten1 (4 096,1) Pool1 (2,2) Conv2 (128,3,3,1) Pool2 (2,2) Fc2 (2 048,1) Fc4 (2 048,1) Conv3 (256,3,3,1) Conv4 (512,3,3,1) Pool3 (2,2) Fc3 F2:(1 024,1) Fc5 F3:(1 024,1) Flatten (2 048,1) Fc1 F1:(1 024,1) Concat F4:(3 072,1) Fc6 F5:(2 048,1) Fc7 F6:(1 024,1) Softmax (8,1) 表 4 StratifiedKFold5折交叉验证下T-CNN的识别效果
% 第N折 N=1 N=2 N=3 N=4 N=5 平均值 识别率 97.92 98.12 95.66 96.38 99.22 97.46 表 5 T-CNN的精确率、召回率以及F1分数
方法 精确率 召回率 F1分数 T-CNN 0.992 3 0.992 1 0.992 2 -
[1] doi: http://pubmedcentralcanada.ca/pmcc/articles/PMC5293267/ CHIANG T Y. Fall Armyworm Migration Across the Lesser Antilles and the Potential for Genetic Exchanges Between North and South American Populations[J]. Plos One, 2017, 12(2):E0171743.1-18. [2] UNBEHEND M, HÄNNIGER S, MEAGHER R L, et al. Pheromonal Divergence between Two Strains of Spodoptera Frugiperda[J]. Journal of Chemical Ecology, 2013, 39(3):364-376. doi: 10.1007/s10886-013-0263-6 [3] NAGOSHI R N, MEAGHER R L. Behavior and Distribution of the Two Fall Armyworm Host Strains in Florida[J]. Florida Entomologist, 2004, 87(4):440-449. doi: 10.1653/0015-4040(2004)087[0440:BADOTT]2.0.CO;2 [4] 张磊, 靳明辉, 张丹丹, 等.入侵云南草地贪夜蛾的分子鉴定[J].植物保护, 2019, 45(2):19-24, 56. doi: http://d.old.wanfangdata.com.cn/Periodical/zwbh201902005 [5] 郭井菲, 赵建周, 何康来, 等.警惕危险性害虫草地贪夜蛾入侵中国[J].植物保护, 2018, 44(6):1-10. doi: http://d.old.wanfangdata.com.cn/Periodical/zwbh201806002 [6] LECUN Y, BENGIO Y, HINTON G. Deep Learning[J]. Nature, 2015, 521(7553):436. doi: 10.1038/nature14539 [7] SCHMIDHUBER J. Deep Learning in Neural Networks:An Overview[J]. Neural Networks, 2015, 61:85-117. doi: 10.1016/j.neunet.2014.09.003 [8] NGIAM J. Multimodal Deep Learning[C]//Proceedings of the 28th International Conference on Machine Learning (ICML-11)[EB/OL]. (2011-06-28)[2019-08-03] http://ai.stanford.edu/~ang/papers/icml11-MultimodalDeepLearning.pdf. [9] doi: http://d.old.wanfangdata.com.cn/Periodical/bjgydxxb201501009 DENG L. Deep Learning:Methods and Applications[J]. Foundations and Trends in Signal Processing, 2014, 7(3/4):197-387. [10] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet Classification with Deep Convolutional Neural Networks[C]//NIPS[EB/OL]. (2012-01-01)[2019-08-03] http://xueshu.baidu.com/usercenter/paper/show?paperid=a5a134e176e1689276201a17412a145a&site=xueshu_se. [11] LU Y Z. Food Image Recognition by Using Convolutional Neural Networks (CNNs)[EB/OL]. (2016-12-03)[2019-08-03] http://xueshu.baidu.com/usercenter/paper/show?paperid=a5a134e176e1689276201a17412a145a&site=xueshu_se. [12] SUZUKI H, MATSUMOTO T, CHUA L O. A CNN Handwritten Character Recognizer[J]. International Journal of Circuit Theory and Applications, 1992, 20(5):601-612. doi: 10.1002/(ISSN)1097-007X [13] 郭井菲, 静大鹏, 太红坤, 等.草地贪夜蛾形态特征及与3种玉米田为害特征和形态相近鳞翅目昆虫的比较[J].植物保护, 2019, 45(2):7-12. doi: http://d.old.wanfangdata.com.cn/Periodical/zwbh201902003 [14] AHMED O. Food Image Recognition by Using Bag-of-SURF Features and HOG Features[C]//International Conference on Human-Agent Interaction[EB/OL]. (2015-10-21)[2019-08-03] http://xueshu.baidu.com/usercenter/paper/show?paperid=a5a134e176e1689276201a17412a145a&site=xueshu_se. [15] DIAZESCOBER J, KOBER V. A Robust HOG-Based Descriptor for Pattern Recognition[C]//Applications of Digital Image Processing XXXIX, SPIE[EB/OL]. (2016-08-26)[2019-08-03] http://xueshu.baidu.com/usercenter/paper/show?paperid=3c8fe9b309a20796eac24747e88963bc&site=xueshu_se. [16] 罗元, 崔叶, 王艳, 等.基于融合DCT和LBP特征提取的面部表情识别[J].半导体光电, 2014, 35(2):330-333, 349. doi: http://d.old.wanfangdata.com.cn/Periodical/bdtgd201402040 [17] doi: http://www.cnki.com.cn/Article/CJFDTotal-MOTO201208009.htm LI X D, YANG W D, DEZERT J. An Airplane Image Target 0 s Multi-Feature Fusion Recognition Method[J].自动化学报, 2012, 38(8):1298-1307. [18] SUN Q S, ZENG S G, LIU Y, et al. A New Method of Feature Fusion and Its Application in Image Recognition[J]. Pattern Recognition, 2005, 38(12):2437-2448. doi: 10.1016/j.patcog.2004.12.013 [19] POISOTAS H. Integrated Management of the Fall Army Worm on Maize: A Guide for Farmer Field Schools in Africa[C]. Rome: Food and Agriculture Organization, 2018. [20] 杜燕飞, 初梓瑞.人工智能助力草地贪夜蛾防控打赢"虫口夺粮"攻坚战[EB/OL]. (2019-7-18)[2019-8-11]. http://country.people.com.cn/n1/2019/0718/c419842-31243174.html.