Research On Detection Algorithm Of Soybean Pests Based On UAV Images

Soybean is one of the important foods and oil crops originated in China.According to domestic statistics,the demand of soybeans in China is increasing year by year.However,affected by extreme weather,all kinds of soybean diseases and pests are developing towards a more harmful and destructive direction,and causing inestimable damage to soybean quality and yield.Therefore,it is very important to grasp the occurrence dynamics of soybean diseases and pests in time and carry out scientific control quickly.Traditional pests detection methods involve observation and judgment by agronomists or experienced farmers at various stages of soybean growth through field work.Soybean is a land-intensive crop,so it is difficult for human eyes to observe soybean diseases and pests in the lush growth period.The traditional methods take a lot of time and labor cost to detect diseases and pests.Based on the principles of image classification and object detection,aiming at the problem of difficult observation of soybean diseases and pests in the field,UAV is applied to soybean diseases and pests detection with the help of the characteristics of flexible flight and simple operation.At the same time,combined with the current mature object detection framework,a soybean diseases and pests object detection algorithm suitable for UAV aerial images is proposed to provide technical support for the precision spraying of pesticide by plant protection UAV.The main work and innovation of this thesis are as follows:(1)In this thesis,soybean diseases and pests are taken as the research object,and leaf phenotype data of soybean during the lush growth period are obtained by UAV aerial photography technology.A total of 3575 images of soybean experimental fields planted in 2020 are collected at a height of 5m above the ground.By means of pixel segmentation and manual annotation with VIA tool,the object detection datasets on aerial soybean leaf phenotype pictures taken by UAV is completed.(2)An algorithm for detecting soybean diseases and pests from UAV aerial images is proposed.The Algorithm uses RetinaNet framework as the basic network.Firstly,Mosaic data argumentation is used to expand the datasets.Secondly,diseases and pests occupy a low proportion of pixels in aerial images.FPN network in the framework is improved to integrate low-level semantic information in the process of feature extraction and retain valuable information of small targets.Finally,GIoU is used to improve the bounding box regression and improve the detection efficiency.The mean average precision(mAP)of the improved network for wormholes and pests reached 0.8393,which is 3.5%higher than that of the original network.(3)In view of the problem that the varieties of soybean diseases and pests in the aerial photos obtained by UAV are not complete,the image classification network GoogLeNet,ResNet-50 and Xception are improved to realize the recognition of soybean diseases and pests by utilizing the characteristics of SDPD with rich varieties of soybean diseases and pests.In this thesis,the method of UAV aerial images detection proposed can greatly improve the detection efficiency of soybean diseases and pests,and has a wide application prospect in the monitoring of field crops.