Application Of High-throughput Object Detectionto Model In Early Diagnosis Of Pine Wilt Disease

Pine Wilt Disease(PWD)is known as the "cancer" of pine trees.In recent years,the occurrence area of PWD has increased year by year,which brings great challenges to the prevention and control work.The early diagnosis of PWD is the key to the prevention and control of PWD,that domestic and foreign researchers have also made outstanding contributions in this field,including the research and development of early diagnosis based on Near-earth Remote Sensing of UAV,which has laid a solid foundation for high-throughput early diagnosis in the forest.However,due to the high requirements of computing power and the authenticity of training samples,there is still a lack of large-scale on-site rapid screening methods,which limits the large-scale application of high-throughput forest early diagnosis technology.In this study,the real image data of early symptoms of PWD in forest were obtained by reverse plant setting,and the early diagnosis model was constructed by combining different machine learning algorithms.The model is evaluated from many aspects,such as model size,operation speed,accuracy and so on.An early diagnosis model of PWD based on real forest data with lower computational power and accuracy was developed.The results are as follows:1.By comparing the hyperspectral(400-1000 nm)information between pine trees infected with pine wood nematode disease and healthy pine trees,it is found that there are obvious spectral differences between healthy pine trees and pine trees infected with pine wood nematode disease in the range of 638-744 nm.In addition,by comparing the gray values of early infection,late infection and healthy pine trees in R,G and B channels,it is found that the gray values of different channels are significantly different,which can be used to distinguish healthy trees from pine trees in early infection and late infection.Therefore,this study collected the image data of the early,middle and late stages of pine wood nematode infection from August to October by means of UAV equipped with hyperspectral imager,and the single sampling interval was 3 ～ 4 weeks.After image preprocessing,the model is established and trained by the image data set collected three times(in which 75% of the data is randomly selected for training the model,and25% of the model is used to verify the accuracy of the model).Annotate the dataset images,namely "AI"(late infected pine)and "PI"(infected early pine),for subsequent modeling.2.The size,accuracy and processing speed of the regular target detection models established by two advanced deep learning models(Faster R-CNN and YOLOv3)and nine conventional machine learning models(k-NN、SVM-linear 、SVM-Gaussian kernel、Gaussian process、decision tree、random forest、MLP-ANN、Ada Boost、Gaussian Na(?)ve Bayes)are compared.the results show that the nine conventional machine learning models can not achieve a balance in size and processing accuracy,and can only be classified from the region of interest.However,the target can not be classified as a whole,so it is not suitable for large-scale identification and diagnosis of PWD.Therefore,we choose the other two deep learning machine algorithms,namely Faster R-CNN framework and YOLOv3 framework,and use these two frameworks to build four models: Faster RCNNRes Net50 、 Faster R-CNNRes Net101 、 YOLOv3 Dark Net53 、YOLOv3Mobile Net.According to the evaluation of the average accuracy,model size and processing speed of the four models,it is concluded that the model of YOLOv3 Mobile Net is small(95.272 MB),the processing speed is the fastest(1.393 samples per second),and the average accuracy(0.632)is only slightly lower than that of YOLOv3 Darknet53,which is the best choice among the four models.3.In order to further explore the application value of the model,this study re-selected a sample plot in the area where pine wood nematodes occurred,in which here are three processing methods in this area,namely,no felling of dead trees(C),preservation of dead trees(T1)and removal of dead trees(T2)to collect UAV image data in this area.Then,four deep school models(Faster R-CNN Res Net50,Faster R-CNN Res Net101,YOLOv3 Dark Net53 and YOLOv3 Mobile Net)were used to predict the infection state(early or late stage of infection)of pine trees under different treatments.The results showed that the number of Masson pine(Pinus massoniana)infected with pine wood nematode disease in the second year was less than that in C treatment regardless of T1 or T2 treatment.Compared with T2,T1 treatment reduced the number of dead trees in the same year,but there were more early infected trees in the following year.In order to verify the accuracy of the prediction results of the model,15 Masson pine samples identified by the model as late infection and 30 early infection were randomly selected and tested by Behrman separation method and PCR test.The results showed that all 45 samples were positive,which proved that the prediction result of the model was accurate.To sum up,this study developed a high-throughput target detection system based on ordinary UAV images and using depth learning algorithm model(YOLOv3 Mobile Net)for early diagnosis of PWD in forest.The system has the characteristics of low computing power,fast processing speed and high detection rate.So,it is suitable for popularization and application in grass-roots forestry system.