| Citrus is one of the most important fruit trees in South China.Citrus industry has become an indispensable part of agricultural economy in South China.Huanglongbing(HLB)is considered to be the cancer of citrus,which spreads quickly and harmfully.The citrus trees infected with HLB will be seriously affected or even die,which will cause huge economic losses to related industries.At present,no drugs are available to cure HLB.Therefore,early detection and isolation must be executed to prevent widely spread of HLB.Experts and special experimental instruments are necessary in the common diagnostic methods,such as PCR detection,DNA probe hybridization and serological diagnosis,thus it cost highly and is low efficiency.As the development of UAV remote sensing technology and the rise of artificial intelligence recently,large amounts of researches on HLB detection using information technologies emerge as the times require.In order to explore the feasibility of UAV remote sensing and artificial intelligence technology in HLB detection,a series of researches were carried out in this study by collecting and analyzing the low altitude multispectral images in the field and the cellphone images on the ground.The main work of this paper includes:1.A method of feature selection based on UAV multispectral image was proposed.ROI was extracted by ENVI software,data augmentation was executed by random averaging,vegetation index was calculated and feature compression was carried out.PCA linear feature and Autoencoder nonlinear feature was compared:and the result showed that linear feature was slightly better than nonlinear feature in HLB detection.2.Some machine learning algorithms such as support vector machine(SVM),K-nearest neighbor(KNN),logistic regression(LR),Navie Bayes and Ensemble learning were compared with each other in HLB modelling.Classifiers were built to identify the multi spectral features of HLB by adjusting parameters.AdaBoost and neural network got the best result to distinguish the healthy plant and HLB-infected one among these algorithms,the accuracy was 100%and 97.28%respectively.3.A method of HLB detection based on cellphone images was proposed.Data augmentation was carried out by using the supervised and unsupervised data augmentation.The efficiency and feasibility between Faster-RCNN model based on the server and MobileNet model based on the mobile client were compared.4.A HLB detection system based on Faster-RCNN model was built.The server supported a large number of clients to run concurrently.It was able to receive and store the pictures from the user and invoke image recognition algorithm and return the classification result to the user.The research results of this paper showed the feasibility of ground-space combined multi-source information in the monitoring of HLB.Through the combination of UAV multispectral remote sensing and machine learning models,it can accurately detect HLB at low altitudes after a series of feature engineering and algorithm tuning optimization.While the initial HLB infected position was confirmed,it was viable to detect HLB further by ground verification on mobile phone and dig the infected plant out to control the HLB spreading. |
