基于深度学习的栽培苜蓿害虫识别模型  

Pest recognition model of cultivated alfalfa based on deep learning

作　　者：张忠雪 冯琦胜[1] 李仲贤 李云昊 王瑞泾 张轩凡 李彦忠[1] 梁天刚[1] ZHANG Zhongxue;FENG Qisheng;LI Zhongxian;LI Yunhao;WANG Ruijing;ZHANG Xuanfan;LI Yanzhong;LIANG Tiangang(State Key Laboratory of Herbage Improvement and Grassland Agro-ecosystems/College of Pastoral Agriculture Science and Technology,Lanzhou University,Lanzhou 730000,Gansu,China;Library,Lanzhou University,Lanzhou 730000,Gansu,China)

机构地区：[1]草种创新与草地农业生态系统全国重点实验室/兰州大学草地农业科技学院,甘肃兰州730000 [2]兰州大学图书馆,甘肃兰州730000

出　　处：《草业科学》2024年第6期1519-1532,共14页Pratacultural Science

基　　金：中国工程院战略研究与咨询项目(2022-HZ-09、2022-XY-139、2021-HZ-5);财政部和农业农村部:国家现代农业产业技术体系项目(CARS-34);甘肃省林业和草原局科技创新项目(kjcx2022010);2023年提前批中央财政林业改革发展资金草原科技支撑项目(2023211)。

摘　　要：苜蓿(Medicago sativa)是我国发展畜牧业的重要优质牧草之一,而病虫害是影响其生长和品质的主要原因,因此准确识别病虫害对其生长发育具有重要意义。YOLO(You Only Look Once)等单阶段目标检测算法通过端对端进行目标检测,RCNN(Region Convolutional Neural Network)等双阶段目标检测算法生成候选区域进行特征提取。为有效识别苜蓿害虫,本文基于YOLOv5和Faster-RCNN两种算法对常见的6类苜蓿害虫进行特征识别,根据召回率(R)、精度(P)、平均精度(mAP)、F1值4种评价指标确定苜蓿害虫识别的最优算法和模型。其中R为样本中的正例被正确预测的比例,F1值为R和P的加权平均值。结果表明:YOLOv5算法识别苜蓿害虫的表现优于Faster-RCNN算法,YOLOv5算法的测试集和验证集精度均高于Faster-RCNN算法。研究结果为苜蓿害虫识别的算法选择提供了科学与理论支撑,对栽培苜蓿管理具有重要意义。Alfalfa(Medicago sativa)is a high-quality forb that is important for the development of animal husbandry in China.Pests and diseases are the main factors affecting alfalfa growth and quality;Therefore,accurate identification of these pests is crucial to alfalfa cultivation.One-stage object detection algorithms such as YOLO(You Only Look Once)performs target detection from end to end and the two-stage target detection algorithm based on RCNN(Region Convolutional Neural Network)generates candidate regions for feature extraction.To effectively identify alfalfa pests,YOLOv5 and Faster-RCNN are used to identify six common pests of alfalfa based on feature recognition in this study.The optimal algorithm and model for identifying alfalfa pests are determined based on four evaluation metrics:Recall,Precision,mAP,and F1 score.Recall is the proportion of positive cases in the sample that are correctly predicted,and F1 is the weighted average of R and P.The results Recall that YOLOv5 outperformed Faster-RCNN in identifying alfalfa pests,with higher precision scores on both the testing and validation sets.This study provides scientific and theoretical support for the selection of algorithms for identifying alfalfa pests,which has important implications for alfalfa cultivation management.

关 键 词：深度学习 YOLOv5算法 Faster-RCNN算法 苜蓿害虫 

分 类 号：S435.4[农业科学—农业昆虫与害虫防治] TP18[农业科学—植物保护] TP391.41[自动化与计算机技术—控制理论与控制工程]