Study Of Citrus Trees Detection And Yield Estimation Technology Based On Plant Protection Unmanned Aerial Vehicle

In recent years,plant protection unmanned aerial vehicles(UAVs)have been widely used for low altitude pesticide spraying due to their high efficiency and safety,especially in precision agriculture of northern China.However,the planting environment in southern China is primarily hilly and mountainous terrain,and the planting spacing is extremely irregular and highly random,such that the UAV spraying mode is unsuitable to be directly adopted in orchard environment.To improve the effect of UAV precise spraying in mountain citrus orchards,a monocular vision based method was proposed to automatically detect trees and extract the tree information such as tree size and location in complex environment.In order to evaluate the spraying effect,this thesis proposed a new approach for data fusion of the information from ground and air to achieve accuracy estimation for the yield of tree.The main research contents and conclusions can be summarized as follows:1.Citrus tree detection based on monocular machine vision in orchard environment.The RGB(Red/green/blue)images were acquired by a UAV surveying on a mountain citrus orchard.In order to reduce the brightness interference on fruit tree detection,the method of selective regions intensity histogram equalization(SRIHE)was developed to improve the region brightness of fruit trees.Because there are significant different color histogram between trees and weed region,the potential fruit tree regions(Regions of interest,RoI)were segmented with the color chromatic mapping method.The color space transformation method(CST)was used to extract 14 color features and the gray level co-occurrence matrix(GLCM)and local binary pattern(LBP)were used to extract 6 texture features from fruit tree regions.The support vector machine(SVM)was used for the process of citrus tree segmentation with radius estimation.The results shows that the SRIHE method can adaptively improve the brightness contrast between the foreground and background,while keeping the color information of images unchanged.It ensured the brightness and the accurate of fruit tree segmentation,with an average segmentation accuracy of 83.09%.The citrus trees detection model based on the appearance characteristics extraction of fruit trees and SVM can further inhibit the interference of scene participants in complex orchard environment,with an accuracy of 85.27%.2.The yield estimation of citrus tree based on the fusion of ground and air information.The RGB image data of fruit trees were captured from ground imaging and low altitude remote sensing.The color chromatic and circle Hough transform(CHT)methods were combined to segment the region of interest(ROI)from the ground image.The color features were extracted using CST,the shape features were extracted using calculating geometric parameters(CGP),and the texture features were extract using local binary pattern(LBP)and GLCM.The extracted features are fed to a SVM to detect the fruits.In order to reduce repeated counting of citrus among 6 ground images with different angles,the fruits regions were extracted Speeded-Up Robust Features(SURF)and matched with Random sample consensus(RANCE)method for statistics the quantity of surrounding fruits of citrus trees.In order to estimate the yield of fruit trees,a multivariate polynomial model was established from the surrounding fruit number and the fruit trees radius.Results show that the proposed method can segment citrus fruit under partial occlusion and is not sensitive to the brightness of the fruit.The SVM detection model can suppress the orchard environmental interferences and reaches a fruits detection accuracy of 96.4%at each ground image.The matching algorithm based on SURF features shows the detecting accuracy of 69.6%for estimating the number of surrounding fruits among 6 ground images from different angles.The accuracy rate of yield estimation was achieved 81.7%using the regression model,whose R~2 is 0.923 and RSME is 78.78.