Research On Pine Pest Identification Based On Drone Images

Pine is the main tree species in forest area.Pine wilt disease(PWD)is a destructive pest caused by pine nematodes,and it is a major forest disaster.Since it was first discovered in Nanjing in 1982,PWD has spread all over the country,resulting in the death of a large number of pine trees and serious damage to the ecological environment of our country.Timely detection of infected pine trees(injured trees)in PWD epidemic areas and taking reasonable control measures is an important means to prevent the further spread of the epidemic.In this thesis,based on the image of PWD epidemic area obtained by unmanned aerial vehicle(UAV),the identification algorithm of PWD is established based on deep learning.The specific research contents are as follows:(1)Identification method of PWD injured trees with different heights based on improved Faster R-CNNDue to the complex and diverse environment of forest areas,the characteristics and dimensions of injured trees in images taken by UAV at different heights have difference.An UAV was used to obtain epidemic sample images at three different heights: 60 m,100m and 220 m.Multi-size PWD injured trees data sets are constructed by using data enhancement methods such as rotation,scaling,adding Gaussian noise and simulating light intensity.The identification method of PWD injured trees with different heights based on improved Faster R-CNN network was proposed.The backbone network improvement strategy,feature bi-directional aggregation strategy and receptive field enhancement strategy are designed respectively to improve the feature extraction ability,the fusion of different levels of feature information ability and target information perception ability of Faster R-CNN network for injured trees captured by UAV at different heights.The experimental results show that based on the improved Faster R-CNN identification model of PWD of different heights,compared with Faster R-CNN identification model and Yolov5 s identification model,the identification accuracy is improved by 37.5% and 4.4% respectively while ensuring the identification stability.(2)Identification method of PWD injured trees with small target based on improved NanoDet-PlusDue to the size of the injured trees in the image taken by UAV at higher height is small,it is easy to be disturbed by the background environment and the detection accuracy is low.Using the image of the epidemic area taken by UAV at the height of220 m,the small-target data set was constructed after data enhancement,and the identification method of small target PWD injured trees based on improved NanoDet-Plus network was studied.The backbone network improvement strategy and feature fusion improvement strategy are designed to improve the feature expression ability of lightweight network NanoDet-Plus to small target injured trees.The experimental results show that the improved NanoDet-Plsus small target PWD identification model is compared with Faster R-CNN,YOLOv4,YOLOV5 s,YOLOV6s,NanoDet and NanoDet-Plus identification models.The model identification accuracy has been improved by 64.9%,21.7%,9.6%,5.9%,1.3% and 1.2%,respectively.The stability of the model increased by 49.5%,30.7%,23.5%,14.9%,4.4% and 2.6%,respectively.The model detection speed is 2.1 times,35.5%,2.8% and 32.4% higher than that of Faster R-CNN,YOLOv4,YOLOV5 s,YOLOV6s identification models.Compared with the current common target detection network,the improved NanoDet-Plus network has significant advantages in detection accuracy,stability and speed.